Single-cell RNA sequencing reads out the gene activity of individual cells, but it destroys each cell it measures and is expensive. Time-course studies of development, disease, or drug response therefore usually have single-cell data at only a handful of time points. Bulk RNA sequencing, which grinds up the whole tissue and reports one averaged profile, is cheap enough to run at many more time points. Current practice uses these two data types separately. Trajectory methods interpolate between the sparse single-cell snapshots and ignore the bulk samples. Deconvolution methods, which estimate which cell types make up a bulk sample, treat each bulk sample as a static puzzle solved against a fixed reference. This note asks a narrower question. Can the bulk measurements taken between single-cell snapshots act as constraints that correct how we reconstruct the cell population at times where no single cell was sequenced?
The core idea fits in two sentences. We model the cell population as a set of simulated cells moving through a low-dimensional state space over time, and we decode each one into gene counts with an overdispersed count likelihood. That avoids the log(x+1) step that distorts sparse data. At time points with only bulk data, we require that the summed decoded counts of the simulated population reproduce the observed bulk composition. Between single-cell snapshots, the bulk data then tells the dynamics where the population actually went, rather than letting it take the shortest path.
Core Claim
Our claim is that a bulk-mean constraint, placed inside a count-valued population dynamics model, measurably reduces error in reconstructing the single-cell distribution at unsequenced intermediate times. It should do better than two alternatives: interpolating between snapshots alone, and deconvolving each bulk sample independently. We expect the benefit to concentrate at branching or rapid-transition windows, where shortest-path interpolation is most wrong. The claim is testable because the right answer is known on existing time-course single-cell datasets. We hide the cells at some time points, give the model only their summed pseudo-bulk, and score how well it recovers the hidden cells.
The novelty is positioned against three pieces of prior work that each hold part of this. DESP solves a per-gene quadratic program. It demixes bulk measurements along an 8-point EMT (epithelial-to-mesenchymal transition) time course into profiles for each cell state. It motivates this work by showing that transitional states carry distinct signatures, but it assumes discrete, pre-clustered states and linear continuous mixing. Count Bridges builds an integer-exact generative bridge from Poisson birth-death dynamics and deconvolves bulk and spatial spots into single-cell count profiles. It outperforms CIBERSORTx, MuSiC and STDeconvolve. It supplies the lesson that count structure matters, but it works on static snapshots with no notion of time. PACE reconstructs continuous trajectories from destructive snapshot time courses. It uses a manifold-aware metric plus transport and bridge fitting. It is our method template and strongest baseline, but it needs dense pre-reduced embeddings, assumes no cell birth or death, and never uses bulk data. None of the three uses bulk measurements as a constraint on continuous single-cell dynamics.
Research Program
The model keeps $N$ weighted particles, each a stand-in for a cell. Particle $i$ has a latent state $z_i(t)$ that evolves under a learned velocity field. It also has a weight $w_i(t)$ that can grow or shrink at a state-dependent rate, which gives a crude account of division and death. This addresses the closed-population assumption that PACE's authors flag as a limitation. A decoder $\mu(z)$ maps a latent state to expected gene proportions, and observed single-cell counts are scored under a negative binomial likelihood with per-cell library size $\ell_i$. This choice follows the evidence of Ling et al. and Armstrong et al. that pseudo-count transforms distort sparse geometry. On the sim-manno-vm benchmark, a negative binomial dissimilarity reached an ARI (adjusted Rand index, agreement between recovered and true clusters, 1 is perfect) near 0.9, against about 0.2 for log-transformed Euclidean distance.
At single-cell time points, the particle population is matched to observed cells with an unbalanced optimal transport loss computed in the decoder's latent space. At bulk-only time points $t$ with observed bulk count vector $y_t$, we add
$$ \mathcal{L}_{\text{bulk}}(t) = -\log \mathrm{Mult}\!\left(y_t;\ \pi_t\right), \qquad \pi_t = \frac{\sum_i w_i(t)\,\ell_i\,\mu(z_i(t))}{\sum_i w_i(t)\,\ell_i\,\mathbf{1}^\top\mu(z_i(t))}, $$
where $\pi_t$ is the gene composition implied by the simulated population. We use a multinomial because bulk sequencing reports only relative abundances. A learnable per-gene scaling term absorbs systematic differences between bulk and single-cell platforms.
The bulk term alone cannot identify a distribution, since many populations share one mean vector. The research hypothesis is that it becomes informative when paired with a smooth dynamics prior anchored at both neighboring snapshots. In that setting it mainly disambiguates two things: how fast mass moves along competing branches, and whether growth or state change explains a shift in composition. Part of the program is to characterize this identifiability empirically. We will measure how much of the held-out error is removed, and at which kinds of transitions.
Baselines
The naive reference point is a time-weighted mixture of the two neighboring single-cell snapshots, with no model and no bulk. This is what a biologist would do by default. Any method that cannot beat it at the held-out time is not worth using.
The strongest existing trajectory method is PACE. It is run on the same retained snapshots with its standard PCA embedding, and we give it the same number of hyperparameter trials as our model. It sees no bulk, which is the point of the comparison. On embryoid body differentiation, PACE reports an average 23.7% reduction in MMD and Wasserstein distributional error across nine held-out experiments, relative to its baselines, and a 15.4% improvement in alignment with RNA velocity. The file gives these only as relative gains, and we found no reported absolute result on Lacuna for PACE's held-out-time error. We will therefore rerun PACE ourselves. Beating it means lower held-out Wasserstein and MMD error across seeds, with non-overlapping bootstrap intervals, on at least two of the three datasets.
The strongest deconvolution competitor is per-time-point Count Bridges. It is trained on the neighboring snapshots and asked to deconvolve the held-out pseudo-bulk into single-cell profiles. It uses bulk and respects counts, but it has no temporal coupling. A fair run uses the same references and the same compute. We will also run DESP with the trajectory binned into many pseudotime states rather than a few clusters. This is the strongest discrete version of our idea, and we generate held-out cells by sampling its state-specific profiles at its inferred fractions.
Three ablations isolate the ingredients. Removing $\mathcal{L}_{\text{bulk}}$ isolates the key claim. Replacing the negative binomial decoder with log(x+1) and PCA tests whether count modeling matters. Freezing the weights $w_i$ tests whether turnover matters.
Why not just…?
Why not just deconvolve each bulk sample independently and smooth the fractions over time? This recovers cell-type fractions at best, not the within-type shifts that define transitional states. DESP shows such states have their own signatures that a static reference dilutes. The per-time-point Count Bridges baseline is the strongest version of this idea, and we will report it directly. If it matches our model on fraction error but loses on distributional error at transitions, that locates exactly what temporal coupling contributes.
Why not just interpolate the snapshots with a good trajectory method and ignore bulk? This is the no-bulk ablation plus PACE. It is a genuine threat if the snapshots are dense enough that interpolation is already near the truth. We expect this on HSMM myoblast differentiation. We therefore vary how many snapshots are withheld and report the bulk gain as a function of gap size. Our claim survives only if the gain grows with gap size and at branch points.
Why not just bin pseudotime finely and run DESP? This is the most serious alternative, because fine bins approximate a continuum and DESP runs in seconds, demixing over 48,000 genes in under 6 seconds on a laptop. It still assumes additive continuous mixing and fixed bins taken from existing single cells. That means it cannot place mass in states between snapshots that no sequenced cell occupied. If fine-binned DESP matches us on held-out error, we will say so and recommend it as the cheaper tool. The experiment below decides this.
Why not just cluster and do pathway-level lineage inference? Tempora orders clusters by pathway profiles and collection time. It scores 0.78 on HSMM and ties PAGA at 0.5 on cerebellum. Its authors state that it cannot capture fine within-cluster change, and it takes no bulk input, so it answers a coarser question than ours.
Proposed Experiments
The primary benchmark is murine cerebellar development, roughly 55,000 cells across 9 time points from E10 to P14, as used by Tempora. We also use murine cortex development (E11.5 to E17.5) and embryoid body differentiation as used by PACE. At each held-out time we sum all cells into a pseudo-bulk. We then corrupt it with gene-specific multiplicative bias and reduced depth, to mimic real platform differences. We evaluate the hidden cells in a fixed evaluation embedding computed from all real cells and not shared with any model. The metrics are Wasserstein-2 and MMD distance, cell-type fraction error, and recovery of held-out transitional marker genes.
The decisive check is whether removing $\mathcal{L}_{\text{bulk}}$ increases held-out error, and whether the full model beats fine-binned DESP and per-time-point Count Bridges at branching windows. The branching windows we will examine are the P0 to P7 cerebellar expansion and the early cortex lineage splits.
A secondary, real-data test uses the EMT time course from DESP. There we withhold scRNA-seq at intermediate points and use the paired bulk proteomics as the constraint, through a learned gene-to-protein link. This test is qualitative and explicitly weaker, because protein-RNA mismatch confounds it.
A clean negative result would still be informative. That would be bulk giving no gain once snapshots are three or fewer steps apart, which would tell experimentalists that extra bulk time points do not substitute for single-cell sampling.
References
- DESP: Demixing Cell State Profiles from Dynamic Bulk Molecular Measurements
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- PACE: Geometry-Aware Bridge Transport for Single-Cell Trajectory Inference
- Dimension Reduction for High-dimensional Small Counts with KL Divergence
- Applications and Comparison of Dimensionality Reduction Methods for Microbiome Data
- Tempora: Cell trajectory inference using time-series single-cell RNA sequencing data