Spending Sequencing Reads on the Reference: Pilot Allocation for Single-Cell Studies on Draft Genomes

Most single-cell RNA sequencing studies measure gene activity in thousands of individual cells by mapping each read back to a reference genome. For most organisms that reference is a draft: a set of fragments with unsequenced gaps between them. Reads from genes that fall in those gaps are thrown away as "unmapped." The resulting loss is not random noise. It removes specific transcripts, and those transcripts are sometimes the markers that separate one cell type from another. Adding more cells or more reads per cell cannot recover them, because the reads have nowhere to land. Wireless engineers know this structure well. A receiver that works with a wrong estimate of the radio channel suffers an error floor that more transmit power cannot remove. Their answer is to spend part of the resource budget on "pilots," which are known signals sent only to learn the channel.

We propose treating the draft reference as the channel estimate, and a small set of deeply sequenced, full-length "pilot cells" as the pilots. The reads from pilot cells are not used for the biological question directly. They are used to build a sample-specific extension of the reference through local reference-free assembly. The remaining budget goes to many shallowly sequenced "payload" cells, which are mapped to the extended reference. The research question is how to split a fixed sequencing budget between pilot and payload, and whether a simple model can predict that split for a new organism from the quality of its draft assembly alone. Our hypothesis is that the best pilot fraction is much larger than current practice assumes when the draft is poor. This mirrors the wireless result that the optimal pilot share can reach half the resource block.

Core Claim

Our central claim has three parts. First, reference incompleteness imposes a floor on cell-type resolution that grows with the missing fraction of informative transcripts and does not shrink with more payload cells. Second, spending a fraction $\alpha$ of the read budget on pilot cells lowers that floor, at the cost of fewer or shallower payload cells, and this creates an interior optimum. Third, the location of that optimum can be predicted from measurable properties of the draft assembly, so it does not require an expensive empirical sweep for each new organism.

The wireless literature supplies the shape of the argument, not the biology. Optimizing Multi-Cell Massive MIMO for Spectral Efficiency shows that once channel estimation error and pilot contamination are modeled honestly, spectral efficiency is maximized by giving half of each coherence block to pilots. This overturns heuristic ratios that treat estimation as a minor cost. Signal Processing for MIMO-NOMA shows the complementary failure: imperfect channel knowledge leaves residual interference that power cannot remove. We use these two results to motivate the form of our model. We do not claim that genomics obeys the same constants.

The Pilot Analogy Made Precise

Let $R$ be the total read budget. Let $\alpha$ be the fraction spent on pilot cells. Let $c(\alpha)$ be the fraction of cell-type-informative transcript sequence that is reachable after the draft reference is extended with pilot-derived assemblies. The number of payload reads that carry usable signal is then $(1-\alpha)R\,c(\alpha)$. A cell-type resolution error with a floor and a variance term can be written as

$$ E(\alpha) \approx E_{\text{floor}}\big(1-c(\alpha)\big) + \frac{\kappa}{(1-\alpha)R\,c(\alpha)}, $$

where $E_{\text{floor}}$ is an increasing function capturing markers that are simply invisible, and $\kappa$ is a dataset-dependent constant setting how fast sampling noise falls with usable reads. This is the analogue of spectral efficiency written as the fraction of the block left for data times a signal quality that depends on pilot length.

The curve $c(\alpha)$ starts at the draft's own completeness $c_0$ and saturates. How fast it rises is bounded by repeat structure. Optimal assembly for high throughput shotgun sequencing proves that reconstruction fails unless reads bridge all interleaved and triple repeats. That paper assumes noiseless bulk DNA and complete reconstruction. We need only its qualitative consequence: gaps inside long repeats are not closed by pilot depth unless pilot reads are long enough to span them, so $c(\alpha)$ plateaus below 1. The model predicts that the optimal $\alpha$ rises as $c_0$ falls, and that it collapses toward zero when the gaps are mostly unbridgeable repeats.

Research Program

The pilot step reuses an existing method. SPLASH finds short sequence "anchors" whose downstream sequences vary across cells and tests that variation with a closed-form statistic. SPLASH+ adds local assembly into "compactors" and compresses 183 million raw reads into roughly 1.5 million interpretable sequences on Tabula Sapiens. We run this pipeline on pilot cells only. We keep compactors that do not map to the draft, and append them to the reference as extra contigs.

Payload cells are then quantified against the extended reference and clustered. The two SPLASH papers analyze all cells reference-free and do not quantify abundance of identical sequences, a limitation both papers note. Our design uses reference-free discovery only where it is cheap and informative, which is in a few deep, full-length cells. It then returns to count-based quantification for the bulk of the cells.

We will build three things. The first is a simulation harness that degrades a complete reference into a realistic draft. The second is the pilot-then-payload pipeline. The third is a fitted version of the model above, in which $c(\alpha)$ and $\kappa$ are estimated from a small pilot run and then used to predict the optimum.

The key analysis is whether the predicted optimum matches the empirical optimum found by sweeping $\alpha$. We will check this across drafts of different completeness and repeat content. The draft-genome literature motivates the degradation settings. The Brugia malayi draft robustly assembles 71 Mb of a roughly 90 Mb genome, with orphan contigs holding essential genes. Furlong et al. and Podani et al. show that gaps silently break annotation and metabolic inference. That is the "absent versus unsequenced" confusion our pilot step is meant to resolve.

Baselines

The naive reference point maps all cells to the draft reference with the whole budget spent on payload cells ($\alpha=0$), then clusters with a standard pipeline. This is what most non-model-organism studies do. Any proposal that cannot beat it on a moderately incomplete draft has no case.

The strongest downstream clustering method in the file is scDFN. It fuses a dropout-aware autoencoder with a cell-neighbor graph autoencoder. It reports a mean ARI of 0.7518 and mean NMI of 0.8131 across 32 datasets, all with conventional references. We will run scDFN on count matrices from every arm, using the same hyperparameter search budget per arm. This separates gains due to better reads from gains due to a better clusterer. We found no reported result on Lacuna for scDFN or any clustering method under deliberately degraded references, so the $\alpha=0$ scDFN arm defines the number to beat.

The strongest competitor on the reference question itself is fully reference-free analysis of all cells with SPLASH+. It is run on the same total read budget, spread evenly over the same number of cells, with its compactors used as features for clustering. Beating it means higher ARI against expert annotation at equal budget, or equal ARI at lower budget.

A second strong competitor pools all payload reads and assembles them, which is equivalent to building the reference extension from shallow reads without dedicated pilots. This is the ablation that isolates our key ingredient: the pilot cells are deep and full-length.

Finally, a fixed heuristic split of 10% pilot tests whether the model's predicted $\alpha$ matters or whether any modest pilot suffices.

"Beating" a baseline means the pilot design at the model-predicted $\alpha$ improves ARI over the $\alpha=0$ arm by more than the seed-to-seed spread. It must also match the empirically best $\alpha$ within one grid step.

Why not just…?

Why not just align to the best available reference and accept the loss? For human data the loss is not small in the regions that matter. SPLASH+ reports that 86% of centromeric anchors and 14% of their supporting reads fail to align even to the complete T2T reference. For non-model organisms there is no T2T reference at all. The $\alpha=0$ arm measures this loss directly. If it turns out that the missing sequence rarely contains cell-type markers, so that the floor term is negligible, the proposal fails, and the experiment will show it.

Why not just run reference-free analysis on every cell? This is the most serious threat. SPLASH already recovered immune diversity in the unannotated mouse lemur. Our argument is budget-based: reference-free anchor testing is described as sensitive to read depth, and it targets sequence variation rather than abundance. Spreading reads thinly over many cells should therefore starve it. If SPLASH+-on-everything matches our design at equal budget, the pilot framing adds nothing, and we will report that result.

Why not just sweep the split empirically for each study? For a well-funded study that sweep is fine, and it is our ground truth. It requires sequencing the same tissue several times, however, which is the cost a planning model avoids. The model is useful only if its prediction, made from a small pilot, lands near the swept optimum. That is the decisive check below.

Proposed Experiments

The primary testbed is Tabula Sapiens Smart-seq2: 10,326 cells across 19 tissues, with full-length reads and expert cell-type labels, as used in SPLASH+. We will degrade the human reference to completeness levels bracketing the Brugia-like 79%. Degradation will be done either by removing contigs at random or by removing repeat-rich regions preferentially. The two schemes separate bridgeable gaps from unbridgeable ones.

For each draft and each $\alpha$ on a grid from 0 to 0.6, we allocate a fixed read budget. Pilot cells keep full depth. Payload cells are downsampled to shallow depth. We then measure ARI and NMI against annotations, the fraction of payload reads assigned, and recovery of the regulated splicing genes SPLASH+ reports as pan-tissue.

External validation uses mouse lemur data analyzed in SPLASH, where the draft is genuinely incomplete. There, we test recovery of MHC and V(D)J-bearing cell populations.

The decisive checks are three. First, ARI versus $\alpha$ must show an interior maximum on poor drafts. Second, the optimum must shift toward larger $\alpha$ as completeness falls and toward zero in the repeat-dominated degradation. Third, the model fitted from a 5% pilot must predict the swept optimum within one grid step across all drafts. If the first check fails, pilot allocation is unnecessary. If the third fails, the idea reduces to an empirical recommendation rather than a planning tool.