Seeing What Dropout Hides: Observation-Geometry-Aware Decomposition of RNA Velocity

A satellite radar measures ground motion only along its line of sight. It is nearly blind to north–south motion, so geodesists combine passes from different viewing angles and state plainly which motion components the data can and cannot resolve. Single-cell RNA velocity has the same structure, though it is rarely treated that way. RNA velocity estimates each cell's direction of change from the ratio of newly made (unspliced) to mature (spliced) transcripts. In any one cell, dropout (genes that are expressed but not captured, and so read as zero) means only a small, cell-specific subset of genes reports on that velocity. Each cell is therefore observed through its own partial "viewing geometry." Some directions of cellular change are well measured, and others are close to invisible.

We propose to import the geodetic approach. First, write each cell's velocity as the result of a known, cell-specific projection. Second, recover it with a regularized joint inversion over neighboring cells whose observed gene sets differ, in the way ascending and descending radar passes complement each other. Third, carry the per-cell resolvability forward into the Hodge decomposition, which splits a flow field into a gradient part (differentiation toward fates) and a rotational part (cycles such as the cell cycle). The claim is that current pipelines implicitly set unobserved velocity directions to zero, as radar analysts do when they assume north–south motion is negligible. This makes cycling motion leak into the apparent differentiation landscape, and the leakage can be measured and corrected.

Core Claim

Let a cell $i$ have a latent velocity $v_i \in \mathbb{R}^k$ in a $k$-dimensional count-model latent space with gene loadings $W \in \mathbb{R}^{G\times k}$. Only the genes $O_i$ with informative unspliced/spliced counts in that cell constrain $v_i$, through the rows $W_{O_i}$. The information the cell carries about its own velocity is then the matrix

$$N_i = W_{O_i}^\top \Sigma_{O_i}^{-1} W_{O_i},$$

where $\Sigma_{O_i}$ is the per-gene noise covariance implied by a Negative Binomial count model. The eigenvectors of $N_i$ with small eigenvalues are that cell's blind directions, the analogue of the radar's north–south null space. We claim three things. First, the spread of $N_i$ across cells is large and structured in real data, because cycling genes and lineage genes drop out at different rates. Second, standard velocity projection followed by Hodge decomposition biases the recovered gradient and rotational components in proportion to this blindness. Third, a resolvability-weighted joint inversion reduces that bias. It also returns calibrated uncertainty and an explicit "unresolved" label, where current pipelines would draw an arrow.

Relation to Prior Work

Two papers motivate this work by decomposing RNA velocity fields. Hodge Decomposition of Single-Cell RNA Velocity meshes a 2D embedding with discrete exterior calculus and splits velocity into curl-free, divergence-free and harmonic parts. It recovers the cell cycle in ductal cells of the endocrine pancreas. Velde applies the natural Helmholtz–Hodge decomposition to build potential landscapes. Both take pre-computed velocities as given and work in UMAP or t-SNE space. Velde explicitly notes sensitivity to velocity quality and to the embedding. We differ in two ways. We do the decomposition in a count-model latent space rather than a 2D visualization. More importantly, we model how dropout shapes what each cell can reveal, instead of treating input velocities as equally reliable.

The geodetic side supplies the method. The two-orbit inversion used in Wang et al. (2020) and Li et al. (2016) separates vertical from east–west motion only by assuming north–south motion is zero. Chen et al. (2022) shows that this assumption pushes real north–south motion into the other estimated components. That leakage is exactly the failure we hypothesize for velocity pipelines. The multidimensional small-baseline inversion of Qu et al. (2022) jointly inverts heterogeneous, misaligned observations with regularization. We reuse its structure: cells in a neighborhood play the role of different orbits, and a smoothness prior plays the role of the temporal regularizer.

On the transcriptomic side, scNBMF supplies the latent space and noise model. It is a Negative Binomial matrix factorization that avoids log-transform distortion and argues that zero-inflation is often overstated. Other work, including Challenges and emerging directions in single-cell analysis and the sparsity masking in Doloris, treats zeros as a distinct dropout process. We do not settle this tension by fiat. The two views imply different $\Sigma_{O_i}$ and so different blind directions, and we test both.

SERM and the survey's discussion of imputation constrain the design. Smoothing-based imputation blurs transitions and can invent structure. Our method therefore never fills the blind directions with neighbors' values; it only reports them. Spatial alignment work such as JADE is a natural later setting for the same idea, but it is out of scope here.

Research Program

The first component is the per-cell observation geometry. We fit an NB factorization in the style of scNBMF to spliced and unspliced counts to obtain $W$ and gene dispersions. For each cell, we define $O_i$ as the genes whose counts give a non-degenerate velocity residual, and compute $N_i$. From it we derive a resolvability score for each latent direction. This is the analogue of geometric dilution of precision in geodesy. It lets us map where in the manifold, and along which directions, velocity is poorly observed.

The second component is joint inversion. We estimate all $v_i$ together by minimizing a weighted misfit $\sum_i \lVert r_i - W_{O_i} v_i \rVert^2_{\Sigma_{O_i}^{-1}}$ plus a graph-smoothness penalty $\lambda \sum_{(i,j)\in E} \lVert v_i - v_j \rVert^2$ on a k-nearest-neighbor graph. Here $r_i$ are the gene-level velocity residuals and $E$ the graph edges. Neighbors with complementary observed gene sets constrain each other's blind directions. The posterior covariance, meaning the inverse of the regularized normal matrix, gives per-cell uncertainty. Any direction whose posterior variance stays near the prior is labeled unresolved.

The third component is decomposition. We run a discrete Hodge decomposition on the kNN graph in latent space, weighting each edge flow by its inverse posterior variance. We then report gradient, rotational and harmonic energies with uncertainty bands, rather than as point pictures.

Baselines

The naive reference point is the standard pipeline: scVelo-style velocities projected into UMAP, followed by the Hodge decomposition of Hodge Decomposition of Single-Cell RNA Velocity or the nHHD of Velde. We run both with their published defaults on the same count matrices and cell subsets.

The strongest existing competitor is a latent-space version of the same pipeline. It uses standard velocities averaged over neighbor moments, projected into the identical NB latent space we use, then decomposed by the identical graph Hodge operator. This isolates our contribution from the "just avoid UMAP" fix that a reviewer would ask for first.

We also include an imputation-first baseline. It runs MAGIC or scVI on counts before velocity estimation, subsampled where needed given SERM's report that these tools did not finish in three days on 0.6 million cells.

The key ablation keeps everything in our method but replaces $N_i$ with the identity. This amounts to uniform trust across cells and directions and removes the observation geometry.

All methods share preprocessing, gene selection, latent dimension and neighbor graph. Each gets the same hyperparameter grid, selected on held-out simulations.

There are no numerical decomposition scores to beat: no reported result found on Lacuna for quantitative Hodge or Velde accuracy on the endocrine pancreas, dentate gyrus, hematopoiesis or cell-cycle datasets. Both papers report qualitative recoveries only. Beating the baselines therefore means lower component-recovery error and lower leakage on simulations with known ground truth. On real data it means greater stability under controlled read downsampling.

Why not just…?

Why not just impute first, then compute velocity? Imputation fills blind directions with neighborhood averages. SERM reports that smoothing imputation blurs developmental transitions, which here would convert unobserved motion into confident but invented arrows. Our simulations test this directly. If imputation-first matches our leakage and calibration, that would weaken the case for explicit geometry, and we would report it.

Why not just do Hodge decomposition in PCA or latent space instead of UMAP? This removes embedding distortion but leaves the blindness caused by dropout untouched. It is the real threat to this proposal, which is why it is our strongest baseline and why the identity-$N_i$ ablation exists. If uniform weighting in latent space performs as well as geometry weighting, the central claim fails.

Why not just bootstrap cells or genes for uncertainty? Bootstrapping measures sampling variance. It cannot reveal a systematic bias from directions that no resample observes, just as resampling radar pixels never exposes the missing north–south motion. We will compare bootstrap intervals to our posterior intervals on simulations where the true bias is known. Coverage failures of the bootstrap in low-resolvability regions would settle this.

Proposed Experiments

We will build simulations with known latent dynamics: a gradient field toward two fates plus a rotational cell-cycle loop, with a tunable ratio between them. Counts are drawn from NB, and optionally zero-inflated, spliced/unspliced models. Dropout is imposed at the sparsity levels used in STAr (0, 10, 30, 50%), and it is deliberately correlated with gene program so that cycling and lineage genes differ in visibility. Metrics are cosine error of the recovered gradient and rotational components, the leakage fraction (true rotational energy appearing in the estimated gradient component), and calibration of the posterior intervals.

On real data we use the endocrine pancreas, dentate gyrus, hematopoiesis and cell-cycle datasets analyzed by the two Hodge papers. There we test two things. First, whether rotational energy concentrates in cells with independently assigned cycling phase. Second, whether decompositions stay stable when reads are downsampled to 50% and 25%.

The decisive check is a correlation test. If leakage in the baselines grows with per-cell blindness (the smallest eigenvalue of $N_i$) and our inversion flattens that relationship, the geodetic framing is doing real work. If leakage is unrelated to blindness, the framing is wrong. A secondary check compares NB against zero-inflated noise models for $\Sigma_{O_i}$. It asks whether the choice changes which directions are flagged unresolved, which would give a concrete, downstream-relevant answer to the scNBMF zero-inflation dispute.