A sequencing run behaves like a shared network link. Every sample loaded onto the machine competes for a fixed number of reads, and within each sample, reads go to transcripts roughly in proportion to their abundance. That abundance is very skewed: a few housekeeping, mitochondrial and ribosomal transcripts absorb much of the budget, while low-abundance regulators get few reads or none. Caching theory has a close match for this situation. It asks how to deliver skewed-popularity content over a bottleneck when receivers already hold some of it locally. It has a well-known answer: under skewed demand, "store the most popular items" (highest-popularity-first) can be arbitrarily worse than a planned allocation. The optimal plan also has a simple shape: at most three groups of items, which are fully stored, partially stored, or not stored.
This note proposes to carry that result into RNA sequencing (RNA-seq, counting RNA molecules to measure gene expression). The "local cache" here is a pretrained expression model, such as scVI or TABULA, that can already predict part of a cell's profile from the rest. The "delivery" is the reads. The design decision is how much to deplete each gene, meaning how strongly to remove its molecules from the library so its reads go elsewhere. The central claim is that the best depletion plan depends on two things together: a gene's abundance and how well the prior model already predicts it. Current practice depletes by abundance alone, which is the sequencing equivalent of highest-popularity-first caching. We expect the jointly optimal plan to separate genes into three groups. That would give a simple, testable rule that differs from current practice.
Core Claim
Let gene $i$ have relative abundance $p_i$ and retention fraction $d_i\in[0,1]$, where $d_i=0$ means fully depleted. Let $R$ be the per-sample read budget and $\sigma_i^2$ the residual variance of the prior model's prediction of gene $i$'s log-expression from the other genes. After depletion, gene $i$ receives about $R\,p_i d_i/Z$ reads, where $Z=\sum_j p_j d_j$. A Poisson-count approximation, in which the variance of a log count is about one over its expected count, combined with the prior prediction, gives a posterior error for gene $i$ of
$$ e_i(d) \approx \left(\frac{1}{\sigma_i^2} + \frac{R\,p_i d_i}{Z}\right)^{-1}, \qquad \min_{d\in[0,1]^G} \sum_i w_i\, e_i(d), $$
where $w_i$ is an analyst-chosen weight on gene $i$ and $G$ is the number of genes. At the optimum, every gene sits at a boundary ($d_i=1$ or $d_i=0$) or at an interior point set by a single shared multiplier on $Z$. We conjecture that the solution sorts genes by a joint index of $p_i$ and $\sigma_i^2$ into three groups: keep fully, deplete partially, and deplete fully and rely on the model. This mirrors the three-group structure proved for linear-program cache placement in Fundamental Structure of Optimal Cache Placement.
Two predictions follow, and both can be falsified. First, some very abundant genes that the model predicts poorly should be kept. Second, some genes of middling abundance that the model predicts well should be depleted. In both cases the plan departs from abundance-only depletion.
We borrow only the placement-under-skew structure from caching theory. The XOR multicast machinery, which combines several users' requests into one coded transmission, has no counterpart here, and we do not claim one.
Relation to Prior Work
The caching literature supplies both the method and the warning. Coded Caching With Nonuniform Demands shows that highest-popularity-first placement is badly suboptimal on skewed Netflix traces. At cache size $M=600F$, a grouped scheme transmits over 2.7× fewer bits, and it matches the highest-popularity-first load with 9.5× less memory. Fundamental Structure of Optimal Cache Placement and Memory-Rate Tradeoff under Nonuniform Random Demands prove the at-most-three-group structure, which cuts the design problem to a one-dimensional search. Our objective is not their linear program, so we must re-derive the structure rather than inherit it.
Correlation-Aware Distributed Caching and Coded Delivery shows that when files are correlated, a file's effective importance depends on what the others already reveal. It splits content into reference files and files predicted from them, and it explicitly leaves the combination of skewed popularity and correlation open. The term $\sigma_i^2$ plays exactly this role: it is the correlation-aware discount on how much a gene is worth delivering. This proposal addresses that open combination in a setting where the correlation structure, gene co-expression, is real and can be measured.
On the sequencing side, Coded Shotgun Sequencing brings channel-capacity reasoning to read sampling but assumes reads fall at uniform positions. That assumption is the sequencing version of the uniform popularity assumption. Known coverage biases motivate dropping it (IVT-seq reveals extreme bias in RNA-sequencing; A benchmark for RNA-seq quantification pipelines). We keep the budget-allocation view but replace the uniform-sampling assumption with measured abundance.
For the prior model, scVI and TABULA supply the predictors. scGeneScope constrains the design in two ways. It finds that fit-to-data models such as scVI often beat zero-shot foundation models, and that performance drops on held-out experimental batches. A model trained on other batches is therefore a stale cache, much as time-varying popularity is the open problem in caching. For this reason the model's $\sigma_i^2$ must be estimated out of batch.
Research Program
The work has three parts.
The first part is analysis. We solve the optimization above and prove or refute the three-group threshold structure. We then characterize when abundance-only depletion is near-optimal. We expect this when $\sigma_i^2$ is roughly constant or uncorrelated with $p_i$. We also characterize when it fails, which we expect when well-predicted genes are not the most abundant ones.
The second part is estimation. We fit scVI through scvi-tools and, separately, a TABULA-style tabular model. For each gene we estimate $\sigma_i^2$ as held-out prediction error from the remaining genes. We then compute depletion plans for a range of budgets $R$.
The third part is in-silico evaluation by read reallocation. Starting from full-depth count matrices, we treat observed counts as truth. We then simulate a depleted library at budget $R$ by multinomial resampling with probabilities proportional to $p_i d_i$. Finally, we reconstruct each profile by combining the resampled counts with the model's prediction.
One physical limit has to be stated up front. In droplet single-cell RNA-seq, depletion applied after capture only reallocates reads among molecules that were already captured. Any gain therefore appears only when libraries are not yet saturated in unique molecules. We will report where each dataset sits on its saturation curve, and we will not claim gains beyond that point.
Baselines
The naive reference is uniform sequencing with no depletion at the same read budget $R$. It shows what any depletion plan must improve on.
The strongest existing practice is abundance-only depletion. This removes the top-$k$ most abundant genes, which is how mitochondrial, ribosomal and hemoglobin depletion is done in practice. It is the direct counterpart of highest-popularity-first caching. We will tune $k$ on a validation split with the same search budget as our method's single threshold parameter. No reported result found on Lacuna for any depletion scheme on these datasets. The caching analogue gives the size of effect that would be worth having: 2.7× lower load than highest-popularity-first in Coded Caching With Nonuniform Demands.
The strongest analysis-side competitor is shallow uniform sequencing followed by model-based imputation. It uses the same scVI or TABULA prior as our method but changes nothing about how reads are allocated. It is strong because TABULA reports imputation error below scGPT and Geneformer on PBMC5K, Jurkat and Melanoma, although the file gives no exact figures. This baseline separates the value of reallocating reads from the value of simply having a good prior.
The ablation that isolates our key ingredient is the three-group rule computed with $\sigma_i^2$ set to a constant, so that allocation depends on abundance alone. If the full method does not beat this ablation, then predictability is not what drives the gain.
All methods share the same datasets, the same prior model weights, the same $R$, and the same downstream classifiers. Beating a baseline means lower weighted error and better downstream metrics at equal $R$. It also means reaching a baseline's accuracy with less $R$, reported as a depth-saving factor in the same form as the 9.5× memory figure.
Why not just…?
Why not just deplete the most abundant genes, as labs already do? This is the real threat. If $\sigma_i^2$ is roughly flat across genes, or tracks abundance, our plan collapses to top-$k$ depletion and the proposal adds nothing. The constant-$\sigma^2$ ablation settles this directly. We will also report the rank correlation between $p_i$ and $\sigma_i^2$ on each dataset before running anything downstream, since the theory predicts gains only when that correlation is weak.
Why not just sequence shallowly and impute afterward? Imputation can only sharpen genes that received some reads or that are predictable. It cannot move reads away from genes the model already knows. The shallow-plus-imputation baseline measures exactly this. If it matches our method, the right conclusion is that a good prior is sufficient and the allocation step is unnecessary.
Why not just buy more depth? Depth has a cost, and past saturation it buys nothing. We report results as equivalent-depth savings, so a lab can compare our plan directly against the price of more sequencing.
Why not skip caching theory and use generic optimal experimental design? The objective is indeed a standard design problem. What caching theory adds is the prediction that the optimum has a three-group shape and that abundance-only placement can fail badly. If the derivation shows no such structure, we will say so and present the result as plain design.
Proposed Experiments
We will use three datasets. scGeneScope provides 627,704 cells across 28 chemical treatments. Following the paper's warning about its confounding batch effect, we exclude Round 2 from any train/test split. We will also use the TABULA benchmark sets: hPancreas for cell-type annotation, Replogle for perturbation response, and PBMC5K for imputation.
We will measure four things. The first is weighted per-gene error against full-depth truth. The second is recovery of differentially expressed genes, meaning genes whose expression changes significantly between control and treatment. We will call these on scGeneScope pseudobulk (cells summed per treatment) at budget $R$ and compare them with the full-depth calls. The third is treatment-identification balanced accuracy on both within-experiment and held-out-experiment splits. The fourth is hPancreas annotation accuracy and Replogle Pearson correlation.
We anchor these against reported numbers: TABULA's 0.9810 hPancreas accuracy, TABULA's +0.069 Pearson margin over scGPT on Replogle, and scGeneScope's 5–7% balanced-accuracy loss for imaging foundation models on held-out experiments. These show how much depletion can degrade results before a method becomes unusable.
The decisive checks are three. The first is whether the full method beats the constant-$\sigma^2$ ablation at matched $R$ on held-out-experiment data, not only within-experiment data. The second is whether the fitted plans actually show the predicted departures from abundance-only depletion: abundant but poorly predicted genes kept, and mid-abundance well-predicted genes depleted. The third is whether the gains survive when $\sigma_i^2$ comes from a model trained on a different experiment, which tests whether a stale cache still helps.
A negative answer on the first check is itself useful. It would show that abundance-only depletion is already near-optimal, and it would give caching-style bounds on how far it can be from optimal.
References
- Bayesian Inference for a Generative Model of Transcriptome Profiles from Single-cell RNA Sequencing
- Tabula: A Tabular Self-Supervised Foundation Model for Single-Cell Transcriptomics
- Fundamental Structure of Optimal Cache Placement for Coded Caching With Nonuniform Demands
- Coded Caching With Nonuniform Demands
- Memory-Rate Tradeoff for Caching With Uncoded Placement Under Nonuniform Random Demands
- Correlation-aware distributed caching and coded delivery
- Coded Shotgun Sequencing
- IVT-seq reveals extreme bias in RNA-sequencing
- A benchmark for RNA-seq quantification pipelines
- scGeneScope: A Treatment-Matched Single Cell Imaging and Transcriptomics Dataset and Benchmark for Treatment Response Modeling
- scvi-tools: a library for deep probabilistic analysis of single-cell omics data