Separating Cell Turnover from Cell Rewiring in Time-Course Bulk Transcriptomes

When a tissue under stress turns up its glycolysis genes, there are two very different explanations. Either the mix of cells changed, for example because more glycolytic immune cells moved in or a glycolytic subclone expanded, or the same cells rewired their metabolism in place. Bulk RNA sequencing measures the average over millions of cells, so it cannot tell these apart. Existing deconvolution tools split bulk samples into cell-type proportions by assuming each cell type has one fixed expression profile. That assumption fails in exactly the time courses where metabolic reprogramming happens, such as hypoxia, drug exposure, and differentiation.

We propose a time-aware deconvolution model. It treats each cell type's expression profile as a point moving along a smooth path in a small learned "latent space", meaning a compressed coordinate system for expression states. Single-cell reference data are typically available at only a few time points, and they pin the path down there. Between those anchors, the dense bulk time series fills in where each cell type's profile must have gone. The model returns two outputs at every time point: cell proportions, and imputed per-cell-type profiles for times with no single-cell data. From these, any bulk change in a metabolic gene can be split exactly into a composition term and a state term. The central claim is narrow and testable. When single-cell references are sparse in time, bulk time-course data carry enough information to recover drift in cell-type profiles, and this recovery beats both static references and naive interpolation of references between time points.

Core Claim and Formulation

Let $y(t) \in \mathbb{R}^G$ be the bulk expression of $G$ genes at time $t$. Let $p_k(t)$ be the fraction of cell type $k$, with the fractions summing to one. Each cell type has a low-dimensional state $z_k(t)$, which a decoder $D_k$ maps to a mean expression profile $s_k(z_k(t)) \in \mathbb{R}^G_{\ge 0}$. We model

$$y(t) \approx \mathrm{diag}(a)\sum_{k} p_k(t)\, s_k(z_k(t)), \qquad \frac{dz_k}{dt} = f_k(z_k, t),$$

where $a$ is a gene-specific platform correction of the kind Bisque learns to fix mismatches between single-cell and bulk capture, and $f_k$ is a small neural network giving the velocity of cell type $k$'s state. At the anchor times $t_j$, where single-cell data exist, $z_k(t_j)$ is fitted so that it encodes the observed cells of type $k$. At all other times $z_k(t)$ is an imputed latent. The only constraint on it is that it follows the ODE and explains the bulk data. The state dimension is kept small, around 2 to 8 per type, so that the drift cannot absorb everything.

Given a fitted model, the change in gene $g$ between two times splits exactly into three parts. The composition term is $\sum_k \Delta p_k\, s_{kg}$. The state term is $\sum_k p_k \Delta s_{kg}$. The interaction term is $\sum_k \Delta p_k \Delta s_{kg}$. Summed over gene sets such as glycolysis, OXPHOS, or the TCA cycle, this yields the scientific output we care about: what fraction of an observed metabolic shift is cell turnover and what fraction is rewiring within a lineage.

The main risk is identifiability. A rising gene could be fit by a growing proportion or by a drifting profile. Four ingredients make the split recoverable. Single-cell anchors fix the profiles at some times. Proportions are shared across all genes while state drift is low-rank per type. Both quantities must change smoothly. And marker genes that stay stable within a lineage, which we estimate from the anchors, tie down the proportions. Whether these ingredients are enough is an empirical question, and synthetic tests with known ground truth answer it before any biological claim is made.

Relation to Prior Work

Static deconvolution supplies the observation model, and its failure under drift motivates this work. Bisque solves non-negative least squares with a sum-to-one constraint, using single-cell references corrected per gene. It shows that mismatch between reference and bulk is the main source of error. But its profile matrix is fixed across samples. Count Bridges generates integer single-cell profiles that sum exactly to a bulk or spatial count. It handles each mixture independently and has no link across time. POPPY aligns bulk samples with single-cell foundation-model embeddings by contrastive learning, which pulls matched bulk and single-cell embeddings together. It is purely cross-sectional. We reuse Bisque's gene-specific correction. Unlike all three, we let profiles move over time and share information across time points.

Dynamic models supply the machinery for latent trajectories, but each one needs single-cell data at every time step. SNODEP fits neural ODE processes that predict future metabolic flux estimated by scFEA, a tool that infers flux from gene expression. FLeCS fits ODEs constrained by gene regulatory networks to single-cell time courses. Neither can take bulk data as input. We borrow their ODE-on-state idea but drive it mostly with bulk observations.

Multi-view latent imputation supplies the logic of using fully observed "anchor" samples to place partially observed ones. DIMVC-HIA and Diffmv both assume views go missing at random, which they handle with masks or reweighting. Here the missingness is structured: single-cell views are absent at whole time points because cells are destroyed when measured. Bulk is the view that is always present.

On the biology side, Montellier et al. found circadian rewiring of liver metabolism under intermittent hypoxia. They worked on bulk liver homogenate and explicitly cannot say whether the change happens across all hepatocytes or reflects shifting subpopulations. This is the kind of question our decomposition is built to answer.

Research Program

We will build the model in three stages. The first stage is a fully continuous version: log-normal or negative-binomial bulk likelihood, per-type decoders pretrained on the anchor single cells, and ODEs solved with the adjoint method as in FLeCS. The second stage adds a hierarchical prior over replicate animals or patients, so that trajectories are partly shared across individuals. The third stage makes the decomposition the reported output, with uncertainty from an ensemble of fits started from different random initializations.

Before any real data, we will map out identifiability on synthetic mixtures. In these mixtures we control the anchor density, the drift magnitude, the rank of the drift, and how strongly proportion changes correlate with drift. The deliverable is a map showing when the composition/state split can be recovered and when it cannot. We will report this map in full, including the region where the method fails.

Baselines

The naive reference point is static Bisque. It uses one pooled single-cell reference and is run separately at each bulk time point, followed by spline smoothing of the proportions. On static data it is strong: Bisque reports Pearson $R = 0.923 \pm 0.064$, against $0.687 \pm 0.450$ for CIBERSORTx, on subcutaneous adipose tissue with 106 bulk samples. This is the number to keep in mind. No reported result was found on Lacuna for any deconvolution method on longitudinal bulk data with drifting profiles.

The strongest competitor is interpolated-reference Bisque. It uses time-matched references at the anchors, interpolates the profiles linearly in expression space between anchors, and deconvolves each time point against the interpolated matrix. It is the strongest because it already captures smooth drift, needs no new model, and differs from our method in only one respect: it ignores what the bulk says about drift between anchors.

We will also run Count Bridges separately at each time point with time-matched references, as the strongest per-mixture generative method. The cited work reports that it beats CIBERSORTx, MuSiC, and STDeconvolve, but the file gives no numbers.

For fairness, every method sees the same bulk samples, the same anchor cells, and the same gene set. Hyperparameters for every method are tuned on an equal budget of 50 configurations, using a validation anchor that is held out from all methods.

Two ablations isolate the key ingredient. One fits the same latent ODE to the anchors only, with no bulk likelihood. This tests whether bulk data add information about drift. The other freezes $z_k$ at the nearest anchor, which tests whether drift modeling matters at all.

Beating the field means two things on held-out anchor times. First, lower error than interpolated-reference Bisque on held-out cell-type profiles and on proportions. Second, a composition/state split for metabolic gene sets that recovers the ground-truth split, with confidence intervals that do not overlap the baseline's.

Why not just…?

Why not just run static deconvolution and smooth the proportions over time? When profiles drift, a static reference forces the drift into the proportions. The method would therefore report turnover where rewiring actually occurred. Bisque's own analysis shows that errors are driven by reference mismatch. The static baseline above measures directly how large this misattribution is on data with known drift.

Why not just interpolate the references between anchors? This is the real threat. If drift between anchors is close to linear, interpolation may match our model. We change the claim to fit this: our advantage should appear only when drift between anchors is nonlinear, for example transient early responses, circadian oscillations, or anchors spaced far apart. The synthetic map and the held-out-anchor experiments will show whether such a regime exists in real data. If it does not, we will report that interpolation suffices.

Why not just run scFEA or SNODEP on the single cells? Those tools require single cells at the time points of interest, and the lack of such cells is exactly the setting here. Feeding scFEA flux scores computed from our imputed profiles is a downstream use of our method, not a replacement for it.

Why not just use Count Bridges at each time point? It enforces exact count sums but gives no way to estimate profiles at times without a reference. With time-matched references it is a baseline; without them it reduces to the static setting.

Proposed Experiments

The first experiment uses pseudo-bulk built from the 16-day human stem-cell differentiation time course of Ori et al., which SNODEP used. We sum the cells at each sampled day into synthetic bulk samples with Dirichlet-varied proportions. We keep single cells at only two or three days as anchors and evaluate imputed profiles and proportions on the withheld days. Ground truth for the composition/state split comes from the true per-day cell-type means. We evaluate it on the glucose and TCA-cycle gene sets that SNODEP examined.

The second experiment uses the K562 and A549 perturbation time courses over 72 hours from FLeCS. These test fast, drug-induced drift. Here we also vary anchor spacing to locate where interpolation breaks down.

The third application is the bulk liver time series under intermittent hypoxia from Montellier et al.. Here we ask whether the loss of rhythm in OXPHOS and TCA transcripts decomposes mainly into hepatocyte state drift or into changes in immune or zonal composition. This application has no ground truth. Results there are hypotheses for spatial or single-cell validation, not benchmark claims.

The decisive check comes from the synthetic study and the held-out-anchor experiments. If the anchor-only ablation and interpolated-reference Bisque come within the uncertainty of our full model on nonlinear drift, then the bulk data add no information about drift. In that case the central claim is false.