Clustering Single Cells by How Their Reads Flow Through Splice Graphs

Many genes can be cut and rejoined in several ways after transcription, a process called alternative splicing, so one gene yields several different messenger RNAs (isoforms). The pattern of exons a cell includes can separate cell states even when the cell's overall gene expression looks the same. Single-cell RNA sequencing should expose these differences. In practice it rarely does: each cell yields so few reads per gene that most splice junctions in most cells have zero coverage. The best current tool for measuring splicing in single cells, ELLIPSIS, works around this sparsity by pooling reads across cells of the same type. It therefore needs cell-type labels before it can measure splicing, and those labels come from clustering on gene expression. The circularity has a cost. Any cell state defined mainly by splicing has to show up in gene expression first, or it is averaged away. The averaging that blurs cell types in bulk tissue reappears here, at the level of splice patterns, inside a single-cell experiment.

We propose to cluster cells directly on their splicing evidence, with no Ψ matrix (Ψ, "percent spliced in," is the fraction of a gene's transcripts that include a given exon) and no prior labels. For each gene we represent splicing as a splice graph: nodes are exons, edges are observed junctions. For every pair of cells we ask one question. Can both cells' sparse junction reads be explained by a single shared, flow-conserving use of that graph, or do they need two different ones? The resulting pairwise compatibility scores feed a consensus clustering procedure: cluster many times on random subsets of genes and cells, and record how often each pair of cells lands in the same group. The output is a stable grouping of cells together with a soft co-membership matrix. That matrix then replaces hard labels as the pooling weights for splicing quantification, which removes the circular dependency.

Core Claim

A pairwise statistic computed on raw junction counts, constrained by flow conservation and evaluated only on the part of the graph that both cells actually cover, makes splicing usable as a clustering signal at single-cell depth. Consensus over resampled genes turns these noisy pairwise votes into stable cell groups. Its specific promise is that it recovers states that differ in splicing but not in expression, where clustering on expression first, which is ELLIPSIS's own pipeline, cannot succeed by construction.

How the Pieces Fit

Each prior paper plays a distinct role in this design. ELLIPSIS motivates the work and supplies part of the method. Its finding that splicing separates glioblastoma core cells from infiltrating peripheral cells independently of expression is the evidence that the gap matters, and its flow-conserving weighted least-squares quantification is what we reuse downstream. What it leaves open is where its pooling groups come from. The single-cell consensus methods SC3, adaSC3 and the consensus cancer-subtype method for scRNA-seq supply the resampling-and-agreement machinery. All three compute Euclidean or correlation distances on gene tables, and these distances degrade on zero-inflated features such as junction counts. We keep their machinery and replace the distance.

SDhaP supplies the conceptual move. It phases haplotypes by treating reads as nodes, scoring each overlapping pair as agreeing or disagreeing, and solving the resulting correlation clustering problem. We apply the same logic one level up: cells are the nodes, and agreement is judged by splice-graph compatibility rather than shared alleles. SDhaP assumes a known integer number of equally sized groups, and we do not. RefShannon supplies the flow-fitting tools. It recovers isoforms from bulk data by sparse flow decomposition solved as a linear program. We use those tools for compatibility testing, not for assembly.

Fundamental Limits of Multiple Sequence Reconstruction from Substrings constrains the design. It shows that the decomposition of a graph into paths is often not unique. So our statistic compares cells at the level of edge flows, which are identifiable from read counts, and never at the level of assembled transcript paths, which may not be.

The Compatibility Statistic

For gene $g$, let $c_i$ be the vector of junction and exon-body read counts of cell $i$ on the splice graph. Let $\mathcal{F}_g$ be the set of nonnegative edge flows that satisfy conservation, meaning inflow equals outflow at each internal exon. We restrict both cells to the subgraph $S_{ij}$ on which each has at least one read in every alternative branch being compared, so that dropout counts as missing evidence rather than as a disagreement. The score is a likelihood ratio:

$$D^{g}_{ij} = \max_{f_i, f_j \in \mathcal{F}_g} \left[\ell(c_i \mid f_i) + \ell(c_j \mid f_j)\right] - \max_{f \in \mathcal{F}_g,\; s_i, s_j > 0} \left[\ell(c_i \mid s_i f) + \ell(c_j \mid s_j f)\right].$$

Here $\ell$ is a Poisson or negative-binomial log-likelihood restricted to $S_{ij}$, and $s_i, s_j$ are per-cell depth scalings. Under the null hypothesis that the two cells share a splice pattern, $D^g_{ij}$ is asymptotically chi-square distributed, with degrees of freedom equal to the number of independent branch ratios in $S_{ij}$. Each score can therefore be converted to a calibrated p-value, and genes with no shared coverage contribute nothing.

In each consensus round we draw about 80% of the genes that carry alternative events and about 80% of the cells. We sum the signed evidence across genes into a cell-by-cell agreement graph, partition it by correlation clustering (greedy pivot plus local refinement, in the spirit of SDhaP's discrete rounding step), and record which pairs are co-clustered. This yields the usual consensus matrix $M$, and the number of groups is chosen by the flattening of the cumulative distribution of $M$'s entries.

Finally, $M$, optionally blended with expression similarity, replaces the hard intra-cell-type weights in the ELLIPSIS quantification step. Each cell borrows strength in proportion to how reliably it co-clusters with others.

Cost is the main engineering risk. Each $D^g_{ij}$ is a small convex fit. We compute it only for genes with alternative events, only over cell pairs within a nearest-neighbour graph built on expression, and with caching across rounds. We will report wall-clock time against SC3.

Research Program

The work proceeds in three stages. First, we build and calibrate the statistic. Under simulated null data, where two cells share a flow and differ only in depth and dropout, the p-values should be uniform. Second, we build the consensus layer and test whether it recovers splicing-only states in simulation and on real data. Third, we close the loop: measure whether soft pooling with $M$ improves Ψ accuracy over hard expression labels, and whether new groups found in real data correspond to known biology.

Baselines

The naive reference point is Leiden or k-means clustering on log-normalized gene expression. This is what most pipelines do, and ELLIPSIS inherits it. On splicing-only states it should score near chance. If it does not, the simulation is leaking expression signal.

The strongest existing clustering method is SC3 run on gene expression, followed by ELLIPSIS quantification using SC3's labels. This is the published pipeline at its best. We run it with SC3's default ensemble and the same number of resampling rounds and gene fraction as our method. We found no reported SC3 clustering-accuracy figure on Lacuna for these datasets, so it will be run fresh. For quantification, the reported ELLIPSIS numbers on its simulated Smart-seq2 data are an average absolute exon error "generally below 5%," precision of 0.79–0.94 and recall of 0.54–0.74 for differentially spliced exons (ELLIPSIS). We first reproduce these numbers in our setting. Beating the baseline means a higher adjusted Rand index (ARI) on splicing-defined states with no loss on expression-defined states, and higher recall for differentially spliced exons at matched precision.

The third baseline is SC3 run on a per-cell Ψ matrix, with missing values handled by pairwise-complete correlation. It tests whether a plain splicing feature table with a standard distance is enough.

Two ablations isolate the key ingredient. The first replaces the flow-constrained statistic with an unconstrained per-junction binomial test, which removes flow conservation. The second runs our statistic in a single clustering pass with no consensus. All methods share the same reads, alignments, splice graphs, compute budget, and an equal hyperparameter search budget.

Why not just…?

Why not just cluster on gene expression and let ELLIPSIS pool within those clusters? For states that differ in expression, this probably works as well as our method, and we expect parity there. The problem is states that differ mainly in splicing. ELLIPSIS's own glioblastoma result shows that such states exist. The planted splicing-only simulation is designed to show exactly this failure: expression-first clustering should score near chance ARI there.

Why not just compute Ψ per cell and run SC3 on it? This is the most serious alternative. With high per-gene depth, as in some Smart-seq2 libraries, it may match our method. Its weakness is that per-cell Ψ is undefined or extremely noisy when a cell has one or two reads on a gene, and correlation over pairwise-complete entries treats a single read as seriously as a hundred. We will sweep sequencing depth by downsampling. If Ψ-SC3 matches us at every depth, the flow statistic adds no value and we will report that.

Why not just alternate: cluster on expression, quantify Ψ, re-cluster on expression plus Ψ, and repeat? This expectation-maximization-style loop is cheap and plausible. Its weakness is that it is seeded by expression clusters and can converge to them. We will run it with the same compute budget. Its gap to our method on splicing-only states, together with whether it ever leaves its initial partition, settles the question.

Proposed Experiments

The first experiment is simulation. Following the ELLIPSIS Smart-seq2 simulation design with annotated and novel junctions, we plant cell groups that differ only in branch ratios on a chosen fraction of genes, and vary three things: the size of the splicing effect, the fraction of affected genes, and the sequencing depth. Metrics are ARI against the planted groups, calibration of the statistic under the null, and Ψ error, precision and recall after pooling.

The second experiment uses the Darmanis glioblastoma data (GSE84465). Recorded tumor location (core versus periphery) provides external labels. The test is whether our clusters separate neoplastic core from periphery cells better than expression-based clusters do, measured by ARI against location within the neoplastic compartment. We will also check whether the differentially spliced exons we recover overlap the migration-related events ELLIPSIS reports.

The decisive check combines the two. On the real data we will report whether the gain over the strongest baseline survives downsampling to 10X-like depth. A gain that appears only at full Smart-seq2 depth would narrow the claim to deep plate-based protocols. A gain that disappears in the splicing-only simulation would refute it.