Abstract: An underlying question for virtually all single-cell RNA sequencing experiments is how to allocate the limited sequencing budget: deep sequencing of a few cells or shallow sequencing of many cells? Here we present a mathematical framework which reveals that, for estimating many important gene properties, the optimal allocation is to sequence at a depth of around one read per cell per gene. Interestingly, the corresponding optimal estimator is not the widely-used plug-in estimator, but one developed via empirical Bayes.
What Problem Does This Paper Address?
Single-cell RNA sequencing (scRNA-seq) is a powerful but resource-intensive tool. Researchers operating on a fixed budget must decide how to allocate their total sequencing "reads." Deep sequencing reduces technical noise (the "graininess" of the data) for individual cells, while sequencing more cells provides a better statistical picture of the whole population.
Until now, advice on this tradeoff has been largely anecdotal. Some argued shallow sequencing is sufficient for clustering, while others insisted deep sequencing is necessary for accurate gene expression estimation. This paper introduces a unified mathematical framework to find the "sweet spot" that minimizes estimation error for a wide range of biological properties, such as mean expression, variance, and correlations.
Methods
The authors model scRNA-seq as a hierarchical process consisting of biological ground truth (cells sampled from an underlying distribution $P_X$) and technical noise (modeled as Poisson sampling). The core of the study compares two types of estimators:
- Plug-in Estimators: The standard approach that treats observed counts (normalized by depth) as the truth. These fail at low depths because they mistake Poisson sampling noise for biological variation, leading to inflated estimates of the Coefficient of Variation (CV).
- Empirical Bayes (EB) Estimators: These "noise-aware" models use the observed data to learn the parameters of the underlying biological distribution, effectively "subtracting" the expected Poisson noise to recover the ground truth.
The study validates this framework using 10x Genomics datasets (PBMC and mouse brain), smFISH (single-molecule fluorescent in situ hybridization) as a physical ground truth, and benchmarks across Drop-seq and CEL-seq platforms.
Key Insights and Ideas
The "One Read" Rule
For estimating moments (mean, variance), correlations, and distribution shapes, the optimal allocation is roughly one read per cell per gene of interest. Surprisingly, this optimal depth is largely independent of the total budget. Once the ~1 read/gene threshold is reached, increasing the budget should generally be used to sequence more cells rather than increasing depth per cell.
The Bottleneck Gene and Experimental Design
In practice, experiments should be designed around the rarest gene of interest. If a marker gene has a relative abundance of $p$, the optimal sequencing depth is $n_{reads} = 1/p$ reads per cell. This ensures that the gene receives, on average, its "one read" required for optimal estimation.
Estimators Matter More Than Depth
Shallow sequencing only succeeds if paired with the correct mathematical tools. Standard plug-in estimators produce biased results (like inflated "dropout" rates) at low depths. In contrast, EB estimators remain consistent across a wide range of sequencing depths, allowing researchers to recover biological signals that would otherwise vanish in shallow data.
Key Results
- Accuracy Boost: In the PBMC dataset, the authors demonstrated that for the marker gene S100A4, the estimation error could be halved by sequencing 10x more cells at 10x shallower depth compared to the original experiment.
- Consistency: EB estimators showed remarkable stability. When subsampling a deep dataset to just 5% of its original reads, the EB estimator recovered gene modules and correlations that completely vanished when using standard plug-in methods.
- smFISH Match: EB estimates of the CV matched the physical smFISH ground truth much more closely than standard scRNA-seq pipelines, which consistently over-estimated variability.
Limitations and Future Work
While the Poisson assumption is well-validated for UMI-based technologies like 10x Genomics and Drop-seq, it may require adjustments for non-UMI or newer spatial technologies. Additionally, the "one read" rule is a heuristic for average gene properties; very rare transcripts or specific splice-variant analyses might still necessitate deeper sequencing. Finally, the impact of this tradeoff on complex "black-box" machine learning workflows, such as deep manifold learning, remains an open area for exploration.
Summary updated