Phase-Invariant Docking: Separating the Cell Cycle from Lineage When Mapping Tumors to Their Cell of Origin

Tumors often keep a transcriptional memory of the healthy cell type they came from. Researchers estimate this "cell of origin" by projecting tumor cells onto reference atlases of normal development and asking where they land. The projection has a known failure. Cancer cells divide constantly, and the division program (genes such as TOP2A, MKI67, PCNA) is loud enough to pull almost any tumor cell toward whatever dividing progenitor the atlas contains. The usual remedy is to regress cell-cycle scores out linearly, but that also deletes real biology. In SHH medulloblastoma, the aggressive SHH-3 subtype is defined by arrest at the cycling granule cell precursor stage (Developmental basis of SHH medulloblastoma heterogeneity).

Wireless engineers face a problem with the same structure and have a clean answer. A radio channel is a sum of propagation paths. Each path's angle and delay is fixed by walls and scatterers, while its phase rotates rapidly with carrier frequency. Downlink channels can therefore be predicted from uplink measurements, without direct supervision, by learning the frequency-invariant geometry and discarding the rotating phase (Learning the CSI Denoising and Feedback Without Supervision; Indoor 5G 3GPP-like Channel Models for Office and Shopping Mall Environments). We propose to model each cell the same way. A stationary lineage coordinate is the "geometry." A cyclic cell-cycle angle is the "phase." The amplitude of the cycle is allowed to depend on lineage. Tumors are then docked using only the lineage coordinate, which should make docking invariant to where a cell is in the cycle but not blind to whether its lineage cycles.

Core Claim

Where a tumor cell sits in the division cycle is a nuisance variable. How strongly its lineage cycles is a lineage property. Linear regression removes both. Leaving the cycle program in place confounds the two. We expect a generative model with an explicit angular cell-cycle variable and a lineage-dependent cycle amplitude to recover known arrest points more faithfully than either choice, and to resist a specific failure: a differentiated-lineage cell that has re-entered the cycle being mistaken for a progenitor.

The wireless analogy supplies the model structure, not a claimed physical equivalence. A channel at frequency $f$ is $h(f)=\sum_l a_l e^{-j2\pi f\tau_l}$, with invariant path gains $a_l$ and delays $\tau_l$. By analogy, we write a cell's normalized expression as

$$ x = \mu(z) + A(z)\,[\cos\theta,\ \sin\theta]^\top + \varepsilon, $$

where $z$ is a low-dimensional lineage coordinate (developmental identity and pseudotime), $\theta\in[0,2\pi)$ is the cell-cycle angle, $\mu(z)$ is the lineage-specific mean expression, and $A(z)$ is a gene-by-2 loading matrix. $A(z)$ sets how strongly and in which genes that lineage expresses the cycle. Its norm is near zero for postmitotic lineages and large for cycling progenitors. The term $\varepsilon$ is count noise, handled with a negative-binomial likelihood inside a conditional variational autoencoder. Docking a tumor cell means inferring its $z$ while marginalizing over $\theta$.

Relation to Prior Work

The reference-mapping machinery comes from Mapping single-cell data to reference atlases by transfer learning (scArches). That work freezes a reference autoencoder and trains small adapter weights so that new query datasets can be projected into the reference space. We reuse its adapter strategy for batch transfer. Its latent space, however, has no structure separating cycle from lineage, and that absence is exactly the gap we target.

Two papers supply the conceptual template, the idea of learning an invariant latent from unlabeled measurements and discarding the rotating component. Learning the CSI Denoising and Feedback Without Supervision trains only on noisy uplink data and transfers across duplex gaps widened from 120 to 480 MHz. GMM-based Codebook Construction and Feedback Encoding in FDD Systems makes the invariant angles explicit parameters. Our contribution is to give the discarded component an explicit parametric form (an angle with lineage-dependent amplitude) rather than hoping a bottleneck ignores it.

A third wireless result constrains the design. CSI Transfer From Sub-6G to mmWave found that direct full-matrix mapping across distant bands fails and that transfer only works in an intermediate angular representation. This is why we dock in the structured $z$ space rather than regressing tumor expression directly onto atlas expression.

On the biology side, three papers motivate the work and serve as evaluation targets. Developmental basis of SHH medulloblastoma heterogeneity shows that malignant cells in the highly differentiated MBEN variant span the whole granule-neuron trajectory, and that consensus subtypes mark distinct arrest points (SHH-3 at cycling GCPs, SHH-2 at postmigratory neurons). Profiling spatiotemporal gene expression of the developing human spinal cord maps ependymoma stem cells onto embryonic neural progenitors. Gene Expression Signature of Normal Cell-of-Origin Predicts Ovarian Tumor Outcomes shows that lineage origin, independent of the oncogenic hit, predicts metastasis and survival.

Two limitations noted in these sources also constrain the project. The ethical window of the reference atlases is narrow (weeks 5–12). Dedifferentiation can make a mature-origin tumor look progenitor-like. We treat the second limitation as a test case rather than claiming to solve it.

Research Program

We would build the model above on the cerebellar granule-lineage reference and on the prenatal spinal cord atlas. The angle $\theta$ gets a von Mises prior. $A(z)$ is a small network whose output is penalized toward zero unless the data require it, so lineages only "cycle" when the evidence says so. Two identifiability anchors are needed.

The first anchor concerns $\theta$. It is initialized from canonical S/G2M gene scores, but the loadings are learned freely across all genes. This lets the model absorb cycle leakage into non-canonical genes, which gene-list methods miss.

The second anchor concerns $z$. It is encouraged to agree with atlas pseudotime within each lineage, and it is penalized for mutual information with $\theta$. We estimate that mutual information with a small adversarial critic that tries to predict $\theta$ from $z$.

Tumor samples are brought in with scArches-style adapters. For bulk cohorts, we derive phase-marginalized lineage signatures from $\mu(z)$ and score them with the Gene Set Variation Analysis procedure Gold et al. applied to MAGIC.

The analysis asks three questions. Does $z$ stay fixed as $\theta$ varies within a lineage? Does $\|A(z)\|$ separate cycling from postmitotic lineages in the reference? Where do tumor cells land, and how much does that landing depend on the cycle signal?

Baselines

The do-nothing reference is scArches mapping with cell-cycle genes left in. We expect this to show the proliferation pull most clearly.

The standard fix is the same mapping after linear regression of S and G2M scores, as practitioners do in Scanpy or Seurat. A gentler variant regresses only the difference between the two scores. That variant removes phase while keeping overall proliferation, so it is the closest simple competitor to our idea.

A third baseline drops all cycle-associated genes before mapping.

For bulk tumors, the strongest existing method is the GSVA-plus-consensus-clustering pipeline of Developmental basis of SHH medulloblastoma heterogeneity. It reports qualitative subtype-to-stage assignments (SHH-3 to cycling GCP, SHH-2 to postmigratory) rather than an accuracy metric. We found no reported quantitative result on Lacuna for cell-of-origin docking accuracy on any of these cohorts. scArches reports throughput and a failure threshold when more than 25% of reference genes are missing from the query, but no docking accuracy.

All methods receive the same highly variable gene set and atlas splits, and the same latent dimension. All neural models get the same hyperparameter search budget and the same GPU hours.

The key ablation removes the lineage dependence of the loading, so $A(z)=A$ is constant. This is a model with an explicit cycle angle but a shared, lineage-independent cycle program. If that ablation matches the full model, the wireless-inspired structure adds nothing beyond "model the cycle as an angle."

Beating the baselines means two things together. First, better recovery of known arrest points on the tests below, with bootstrap confidence intervals over cells and donors that exclude zero difference. Second, no loss of cross-lineage separation in the reference.

Why not just…?

Why not just regress out the S–G2M difference score? This is the real threat. It already keeps "is this cell cycling" while removing "which phase." It is linear and tied to a fixed gene list, though. If different lineages express the cycle through different gene loadings, a single linear removal will under-correct some lineages and over-correct others. The synthetic re-entry test below decides this directly. If difference regression docks re-entered cells as well as our model does, we would report that and recommend the simpler method.

Why not just drop cell-cycle genes? Removing named genes leaves the downstream programs they drive: metabolic, chromatin, and biosynthesis genes that covary with division. CSI Transfer From Sub-6G to mmWave gives the analogous warning that crude feature-space surgery underperforms a structured intermediate representation. That is an analogy, not evidence, so the experiment has to show it. We will measure how much cycle signal remains in the lineage latent after gene dropping, using the adversarial critic, and compare docking accuracy.

Why not just add cycle score as a covariate in scVI or scArches? A covariate conditions the decoder on the cycle but does not force the lineage latent to be independent of it. It also assumes one cycle effect shared across lineages. This baseline is essentially our constant-$A$ ablation with a scalar in place of an angle, and we run it as such.

Proposed Experiments

The first decisive check is synthetic lineage re-entry. In the spinal cord atlas, several lineages divide (neural progenitors, glial precursors) and others do not. We take held-out postmitotic cells and add the cycle program estimated from a different lineage's cycling cells, at controlled amplitudes. This mimics a differentiated cell re-entering division.

We then measure the fraction of these cells docked to their true postmitotic lineage and the shift in pseudotime. Baselines without regression should drift toward progenitors. Full regression should behave well here but fail the second test. The full model should resist the drift, and the constant-$A$ ablation isolates whether lineage-dependent loading matters. An honest caveat is that this simulation is only as realistic as the transplanted program.

The second check is arrest-point recovery on medulloblastoma. We map the MBEN snRNA-seq tumors onto the granule-lineage reference and test whether the pseudotime distribution of malignant cells still spans precursors to postmigratory neurons. We then score the 223-sample MAGIC SHH cohort with phase-marginalized signatures.

Success means reproducing SHH-3 as cycling-GCP-arrested and SHH-2 as postmigratory, and doing so with less dependence on the cycle than the regression-free baseline. We measure that dependence as the correlation between assigned stage and raw G2M score within subtype.

We also check the spatial data. The spatial immunohistochemistry of MBEN nodules shows proliferating cells at the periphery and VSNL1+ differentiated cells in the center. Docked stage should follow this radial gradient, but cells in different cycle phases within the periphery should receive the same lineage stage.

The third check tests generality on ependymoma. We re-dock ependymoma stem cells onto the spinal cord atlas and ask whether the neural-progenitor docking and the progenitor-like versus astrocyte-like split reported by Li et al. survive when the cycle is marginalized. If the docking weakens, it was partly driven by the cycle. The Merritt ovarian signature serves as an external outcome check where cycle confounding is plausible. On the Tothill cohort, we test whether phase-marginalized fallopian-tube-like classification preserves the reported survival separation.

A result where difference regression ties our model on all three checks would be a useful negative finding. It would mean the cycle-lineage confound is essentially linear and needs no richer model.