Reading the Cell-Cycle Phase Distribution Out of Bulk RNA-seq

When a tumor biopsy is sequenced in bulk, every cell's transcripts are poured into one tube. Existing tools for unmixing that signal, called deconvolution, ask which cell types are present and in what proportions. They are silent on a question that matters just as much for proliferating tissue: where the dividing cells sit along the cell cycle. The cell cycle is a loop, not a set of boxes. A cell's position on it is an angle, a "phase" on a circle, and gene expression changes smoothly as the phase advances. A drug that arrests cells before division and a drug that slows the whole cycle can shift average proliferation markers by the same amount. They leave different phase distributions behind, and a bulk marker score cannot tell them apart.

This proposal borrows the mathematics that structural biologists built for angles on circles and applies it to bulk transcriptomes. Protein dihedral angles live on circles, and there the field replaced binned rotamer libraries with smooth circular distributions. We do the same for cell-cycle phase. The central claim is precise and testable. Bulk expression identifies exactly a small, finite set of low-order circular summaries of the hidden phase distribution, its first few "trigonometric moments". The number of recoverable moments is set by how wiggly gene expression is along the cycle. A mixture of von Mises distributions, the circular analogue of the Gaussian, is the natural low-parameter object to fit to those moments. Fitted jointly with a non-cycling fraction and a count likelihood, it yields a calibrated estimate of the cycling fraction and the phase density from bulk data alone.

Core Claim

Write the bulk mean count of gene $g$ in sample $j$ as

$$\mathbb{E}[Y_{gj}] = s_g N_j \Big[(1-c_j)\,a_g + c_j \int_0^{2\pi} p_j(\theta)\, x_g(\theta)\, d\theta\Big].$$

In this equation $N_j$ is library size and $s_g$ a gene-specific platform scaling between single-cell reference and bulk. $c_j$ is the fraction of cycling cells, and $a_g$ is the expression profile of non-cycling cells. $p_j(\theta)$ is the phase density on the circle, and $x_g(\theta)$ is the expression of gene $g$ at phase $\theta$. Expression along the cycle is smooth, so we represent $x_g(\theta)$ by a short Fourier series of order $L$ (sums of $\cos l\theta$ and $\sin l\theta$ for $l \le L$). The integral then collapses to a linear function of the trigonometric moments $m_{l,j} = \mathbb{E}_{p_j}[e^{il\theta}]$ for $l \le L$. For a von Mises component with mean $\mu$ and concentration $\kappa$, these moments have the closed form

$$m_l = e^{il\mu}\, \frac{I_l(\kappa)}{I_0(\kappa)},$$

where $I_l$ is the modified Bessel function of order $l$. Two consequences follow, and we will test both. First, bulk data constrain $c_j$ and the $2L$ real numbers $c_j m_{l,j}$ and nothing else about $p_j$. Any claim of finer phase resolution from bulk is extrapolation. The practical question is how large $L$ is for real cell-cycle genes, which we will measure rather than assume. Second, a von Mises mixture with few components is identified whenever its parameter count does not exceed the recovered moments. That makes the choice of mixture size an identifiability question with a principled answer, not a tuning knob.

Where This Sits in Prior Work

The deconvolution literature motivates the gap and supplies constraints. Bulk averaging hides stochastic and clonal cell states (Genome 2011). In tamoxifen-resistant breast cancer lines, bulk profiles misread the selection of a stem-like subpopulation as uniform adaptation (Lin et al.). Proliferative state is exactly the kind of heterogeneity at stake in such cases. The current methods share one assumption: cell states are discrete categories. MEAD corrects platform scaling between single-cell references and bulk and gives valid confidence intervals, but only over fixed types (Statistical inference for cell type deconvolution). We reuse its gene-specific scaling term $s_g$ and its sandwich-variance approach, and replace the categorical latent with a circular one. POPPY aligns bulk and single-cell embeddings using distances on the Cell Ontology tree (Cross-modal transfer learning for mapping bulk transcriptomes at cellular level). A tree has no loops, so it cannot represent a cyclic process. Count Bridges argues that count data must be modeled on the integers and deconvolves bulk data into latent single cells with exact sum constraints (Count Bridges enable Modeling and Deconvolving Transcriptomic Data). We take its point about counts by using a negative-binomial likelihood on bulk reads. However, its latent cell identities carry no phase geometry. THUNDER shows that discrete bin-based factorization degrades as bins get finer and sparser (THUNDER). We expect the same failure when the cycle is sliced into many phase bins.

The conformational-sampling literature supplies the method. Bivariate von Mises models on the torus replaced binned Ramachandran statistics without boundary artifacts (A generative, probabilistic model of local protein structure). Side-chain angle distributions can be conditioned on exact real-valued backbone angles rather than grid cells (Beyond rotamers). That is the analogue of our $x_g(\theta)$ conditioned on continuous phase instead of discrete G1/S/G2M labels. Minimum message length inference picks the number of von Mises mixture components automatically (Getting 'ϕψχal' with proteins). We adopt it, capped by the identifiability bound above. VonMisesNet shows that closed-form von Mises mixtures are far cheaper than iterative diffusion over angles, reporting a 47x GPU speedup over Torsional Diffusion (Von Mises Mixture Distributions for Molecular Conformation Generation). This supports using closed-form moments over a generative sampler in the inner loop. For evaluation we use the Wasserstein machinery on the flat torus, specialized to one circle (Two-sample goodness-of-fit tests on the flat torus). None of these works touches mixtures of cells. This proposal is their first use on a heterogeneity-masking problem.

Research Program

The work has three parts. First, we build a phase-resolved reference. We take single-cell data with cycling populations, chiefly malignant cells in the Curated Cancer Cell Atlas (3CA; 2.7M cells, 1,631 samples, used in POPPY). We assign each cycling cell a phase angle by fitting a circle to its cell-cycle gene embedding. We then fit Fourier expression curves $x_g(\theta)$ per gene by penalized negative-binomial regression and fit the non-cycling profile $a_g$. This step measures $L$. It tells us how many Fourier harmonics are needed before held-out single-cell likelihood stops improving, and therefore how many phase moments bulk can ever reveal.

Second, we build the estimator. It fits $c_j$ and a von Mises mixture $p_j$ by maximizing the bulk negative-binomial likelihood, with MEAD-style scaling $s_g$ estimated on genes shared across platforms. The mixture size is chosen by message length under the identifiability cap. Uncertainty comes from sandwich variances on $(c_j, \mu, \kappa)$, propagated to the density by the delta method. We also derive the moment-identifiability result formally, including conditions on the rank of the gene-by-harmonic coefficient matrix. That rank determines whether cycling fraction and phase shape are separable, since separability requires $a_g$ to differ from the cycle-averaged profile.

Third, we run a cohort-level test. We ask whether group differences in phase distributions can be tested with calibrated error rates, using the circular Wasserstein two-sample test on posterior phase densities across donors.

Baselines

The naive reference point is a proliferation score. It averages standardized S-phase and G2/M marker expression in bulk and maps that average linearly to a cycling fraction, with the phase distribution fixed to the reference-population average. Any method that cannot beat this on phase-density error has learned nothing beyond "how much is dividing."

The strongest existing competitor is MEAD run with three discrete phase categories (G1, S, G2/M) plus a non-cycling category as its cell types. It is built from the same reference cells, uses the same genes and the same scaling correction, and gives each method the same hyperparameter search budget. MEAD is the right opponent because it already handles platform scaling and inference validity. That isolates our change: a continuous circular latent versus categories. We convert its output to a step-function density on the circle for comparison. We found no reported result on Lacuna for cell-cycle phase deconvolution from bulk data. MEAD's islet benchmark is summarized only qualitatively in the file, as "competitive to superior" accuracy over CIBERSORT, MuSiC and OLS. Count Bridges likewise reports beating CIBERSORTx and MuSiC on cell-type proportions without a number in the file. We will run Count Bridges with phase-bin labels as a secondary generative competitor if compute allows. Beating MEAD means lower circular Wasserstein error in phase density at equal or better error in cycling fraction. It also means nominal 95% interval coverage for the cycling fraction, both on held-out 3CA patients and on the harder cases below.

Two ablations isolate the key ingredients. The first is fine-binned deconvolution: 24 phase bins, non-negative least squares, and a circular smoothness penalty, with the penalty tuned on validation data. It tests whether the parametric von Mises form adds anything beyond smoothness. The second keeps our model but replaces the von Mises mixture with a Gaussian on unwrapped phase, cut at the M/G1 boundary. It tests whether respecting periodicity matters. That should show up specifically for distributions with mass straddling mitosis.

Why not just…?

Why not just deconvolve G1, S and G2/M as three discrete cell types? This is the MEAD baseline, and on cycling fraction alone it may well match us. Its weakness is structural. Three categories give at most a three-bin histogram, so they cannot distinguish a G1 arrest from uniformly slowed G1 transit within the same phase label. The mixture-design experiment below targets exactly those cases. If discrete phases match us there too, the continuous formulation is not worth its complexity and we will say so.

Why not just use many fine phase bins with a smoothness penalty? This is the real threat, and the moment argument explains why. Fine bins see exactly the same $2L$ moments that our model sees, so any advantage must come from better extrapolation and uncertainty, not from new information. We expect the binned solution to be ill-posed beyond $L$ harmonics, with bin weights that swing with the penalty strength. THUNDER's finding that fine discrete bins degrade under sparsity points the same way. If the tuned binned estimator matches us on density error and coverage, we will reframe the contribution around the identifiability result and the measured $L$. The claim would then be that bulk data carry $2L$ phase moments and no more, whichever estimator one uses. That would still be a useful finding.

Why not just report a proliferation marker score? A score is one number. It is a noisy projection of $c_j$ mixed with $c_j m_{1,j}$, and it cannot separate "fewer cells cycling" from "cells piling up at one phase." The naive baseline quantifies this directly. The decisive check is whether the score confuses matched-score mixtures that our model separates.

Proposed Experiments

The primary benchmark uses pseudo-bulk samples from 3CA malignant cells. We sample cells to hit prescribed pairs of cycling fraction and phase density, then aggregate their counts. The prescribed densities cover four cases: uniform cycling, a concentrated G1-like peak, a G2/M peak, and a density straddling mitosis. We hold out whole patients from reference construction. Metrics are circular Wasserstein-1 distance between true and recovered phase densities, absolute error in cycling fraction, and interval coverage.

The decisive design uses matched-score pairs. These are mixtures constructed to have identical proliferation-marker scores but different phase distributions. Success is correctly telling each pair apart at a controlled false-positive rate, and doing so when discrete three-phase MEAD cannot.

Cross-platform robustness uses the Xin and Segerstolpe pancreatic islet datasets, each used as reference for the other, as in MEAD. Islet endocrine cells rarely divide, so this doubles as a negative control. The method should report a small cycling fraction with wide phase uncertainty, not a confident phantom peak. The Onek1k blood-cell cohort, used by Count Bridges, serves as a second low-proliferation control.

Finally, we measure $L$ on held-out single cells and check the identifiability theory. On synthetic data generated with known $L$, recovery error should collapse for mixtures within the moment budget and plateau for mixtures beyond it. If recovery improves beyond that bound, the theory is wrong and must be revised before any biological claim is made.