Automated tools for single-cell RNA sequencing (measuring which genes are active in thousands of individual cells) now let a bench biologist get a cell-type label for every cell without writing code. The label arrives as a bare name, "CD4+ memory T cell", whether it came from a richly sequenced cell or from one with a few hundred captured molecules. Speech recognition hit the same problem decades ago and developed a remedy called uncertainty decoding. When a noise-cleaning front end hands a recognizer its best guess of the clean signal, it also hands over how unsure it is. The recognizer then adds that uncertainty to its own model variances, so muddy evidence counts for less. We propose to carry this remedy into cell annotation. A probabilistic encoder already reports a per-cell variance, and we would add that variance to the class models of a reference-based annotator. The annotator then backs off to a coarser, still-correct label (for example "T cell") when a cell's measurement is too thin to support a fine one.
The second half of the idea is aimed at non-experts. Because measurement noise enters the decision explicitly, the tool can say why it coarsened a label. "Ambiguous because the measurement is thin" means more sequencing depth would probably resolve the cell. "Ambiguous even with a clean measurement" means the cell sits between reference types. A biologist can act on that distinction without understanding latent spaces. The user sets only one number, the fraction of wrong labels they are willing to accept.
Core Claim
The claim has two parts. First, rejection decisions driven by calibrated measurement uncertainty will make fewer confident errors than rejection driven by the classifier's own output confidence, and the gap will grow as sequencing depth falls. Second, comparing the decision made with and without the measurement variance separates depth-driven from biology-driven ambiguity, and that separation predicts whether deeper sequencing of the same cell would resolve it.
The mechanism comes from the speech literature. Observation uncertainty measures for sparse imputation added estimated reconstruction uncertainty to acoustic-model variances. On babble noise at 10 dB it cut letter error from 21.6% to 18.4%. An oracle uncertainty reached 11.4%, which shows both the value and the difficulty of estimating uncertainty well. Improving RNN Transducer With Target Speaker Extraction and Neural Uncertainty Estimation found that a recognizer fed deterministic enhanced speech produces confident "hallucinated" insertions, and that explicit uncertainty inputs cut these by 17.5%. We expect the analogous failure in annotation: a shallowly sequenced cell whose embedding drifts toward the population average and is confidently assigned to a wrong or generic type.
Research Program
The front end is scVI, from A deep generative model for gene expression profiles from single-cell RNA sequencing. It maps each cell to a Gaussian posterior over a low-dimensional latent state, with mean $\mu$ and diagonal variance $\sigma^2$. As the research file notes, downstream toolkits almost always keep $\mu$ and throw away $\sigma^2$.
On a labeled reference, we fit a Gaussian for each leaf cell type $c$ in the latent space, with mean $m_c$ and covariance $S_c$. For a query cell, the probability of its latent state under each type, integrating over the cell's own uncertainty, has a closed form:
$$p(\text{cell} \mid c) = \mathcal{N}\big(\mu;\; m_c,\; S_c + \mathrm{diag}(\sigma^2)\big),$$
This is exactly the variance-addition rule of classical uncertainty decoding. We obtain leaf posteriors by Bayes' rule with reference class frequencies. The probability of an internal node of the cell-type tree is the sum of its children's probabilities. The tool returns the deepest node whose probability exceeds a threshold $\tau$.
Following the partial-rejection idea of Uncertainty-aware single-cell annotation with a hierarchical reject option, we calibrate $\tau$ on a held-out reference split. It is set as the value at which the error rate on accepted labels matches the user's tolerance. That paper derives uncertainty only from classifier outputs and leaves the thresholds to the user. We change both.
Diagonal rather than isotropic variance is essential here. The ablation in Xi-Vector Embedding for Speaker Recognition showed a 38.7% EER degradation when per-dimension variances were collapsed to a single scalar. This result constrains our design.
Amortized encoder variances are often miscalibrated, so raw $\sigma^2$ may not track the real error. We therefore learn a per-dimension rescaling of $\sigma^2$ from reference cells. Each reference cell is binomially thinned to lower depth. The target for the rescaling is the squared distance between the thinned cell's $\mu$ and the full-depth cell's $\mu$. This mirrors the learned uncertainty transform of Gemmeke et al. and the error-predicting uncertainty network of Shi et al. The same construction gives an oracle variance on test cells, which bounds how much better uncertainty estimation could help.
The reason code comes from running the decision twice. The first run uses $\sigma^2 = 0$ and asks whether the cell would be confidently resolved if perfectly measured. The second run uses the calibrated $\sigma^2$. If the cell is fine-grained under the first run but coarsened under the second, it is flagged "measurement-limited." If it is coarse under both, it is flagged "biologically ambiguous relative to the reference."
For accessibility, the whole procedure ships as one call that takes a reference, a query, and an error tolerance. It is also registered as a tool in the style of the sc-Omni toolkit used by CellAgent, so that a conversational agent or a graphical front end like PIVOT can surface labels and reason codes without parameter tuning.
Relation to Prior Work
The closest work is TISSUE. It calibrates uncertainty for imputed spatial gene expression with conformal methods, then uses that uncertainty mainly for post-hoc cell filtering, multiple imputation, and weighted PCA. We keep every cell and move uncertainty inside the classification likelihood. The output is a coarser label rather than a discard. TISSUE also targets imputation error in spatial data, while we target depth-driven measurement noise in dissociated scRNA-seq.
The hierarchical reject paper supplies our decision structure but, as noted above, treats inputs as perfectly observed. The packet-loss work of Combining packet loss compensation methods for robust distributed speech recognition motivates soft rather than hard handling of degraded observations: continuous confidence weighting beat binary masking under 50% bursty loss.
LLM-based annotators, benchmarked in Single-Cell Omics Arena and used inside CellAgent, work from deterministic marker lists. Our tool gives such agents a per-cell signal they currently lack, but we do not depend on a proprietary model to produce it.
Baselines
The naive reference point is flat annotation with no rejection. A classifier (logistic regression and a small multilayer perceptron) is trained on scVI means and assigns every query cell a leaf label. This shows how many confident errors appear as depth falls.
The strongest existing method is hierarchical partial rejection driven by classifier probabilities, as in Uncertainty-aware single-cell annotation with a hierarchical reject option. We will reimplement both its best flat classifier and its top-down hierarchical classifier on scVI means with highly-variable-gene selection, which that paper found consistently helpful. We found no reported balanced-accuracy result on Lacuna for this method on its benchmark datasets, so we will reproduce it rather than quote a number. It receives the same scVI model, reference split, and threshold-calibration procedure as ours, and the same hyperparameter search budget. Beating it means one of two things at every depth below full: strictly less hierarchical information lost at a matched error rate on accepted labels, or a lower rate of confident errors at matched coverage.
A second strong competitor is Monte Carlo propagation. We draw 50 latent samples per cell from the scVI posterior, classify each with the flexible baseline classifier, and average. This is the "multiple imputation" route used by TISSUE, and it is the most general way to use $\sigma^2$.
Two ablations isolate the key ingredient. The first uses our closed-form decoder with $\sigma^2$ set to zero (reference Gaussians alone), which separates the effect of the class model from the effect of the uncertainty. The second uses raw versus calibrated versus oracle $\sigma^2$, which measures the heuristic gap as in the speech literature. A third ablation feeds $\log\sigma^2$ as extra classifier inputs, the auxiliary-input design of Shi et al.
Why not just…?
Why not just filter out low-quality cells with standard count thresholds? Filtering throws away entire cells, including shallow cells whose coarse identity is perfectly recoverable. The hierarchical reject paper showed full rejection loses substantially more information than partial rejection. We run a baseline that drops query cells below a total-count cutoff, and we compare at matched numbers of correct labels at each tree level. If filtering matches us, the per-dimension variance is adding nothing beyond library size, and we would say so.
Why not just trust the classifier's confidence, since low-depth embeddings should already look ambiguous? This is the real threat. Whether shallow cells land in ambiguous regions or confidently in wrong ones is an empirical question. The confident-error analysis across thinning levels, described below, settles it directly. If classifier entropy rises in step with depth loss, the first half of our claim fails. The reason-code decomposition would remain, because classifier entropy alone cannot say whether more sequencing would help.
Why not just sample from the scVI posterior and vote? For a flexible downstream classifier, sampling can approximate the same marginalization and might beat our Gaussian class model. We run it as a full baseline. If it wins on accuracy, we adopt sampling as the decoder and keep the calibration and decomposition as the contribution, since both apply unchanged. The closed form is preferred only if accuracy ties, because it is deterministic and about 50 times cheaper. That matters for tools run on modest hardware.
Proposed Experiments
We will use the five hierarchically annotated datasets from the hierarchical reject benchmark (AMB, COVID, AzimuthPBMC, Flyhead, Flybody) plus PbmcBench, with the same cell-type trees. For each dataset we split reference from query. Query cells are binomially thinned to 50%, 25%, 10%, and 5% of their original counts, which plays the role of the signal-to-noise sweep in the speech work. Because the full-depth version of every thinned cell is known, we get both ground truth and an oracle.
Metrics fall into four groups. The first is accuracy-rejection curves and retained hierarchical specificity (average depth of accepted labels) at fixed error tolerances of 5% and 10%. The second is the confident-error rate: accepted labels that are wrong. The third is calibration error of node probabilities per depth bin. The fourth covers the decomposition: the AUROC of the "measurement-limited" flag for predicting whether the full-depth copy of the cell resolves to a finer correct label.
The decisive checks are three. First, at 10% depth and below, our decoder must show lower confident-error rates than the probability-based hierarchical reject at matched coverage; otherwise the core claim is false. Second, calibrated variance must beat raw variance, and we will report how much of the oracle gap it closes. Gemmeke et al. warn that at severe noise, heuristic uncertainty stopped helping. We will report honestly if our 5% condition behaves the same way.
Third, the measurement flag must rise monotonically with thinning. On untouched full-depth cells it must stay low, and there the biological-ambiguity flag should concentrate on sibling types that the reference itself confuses.
As a light accessibility check, we will expose the tool through a CellAgent-style tool registry and verify that an open-weight agent can call it from a natural-language request with no parameters beyond the error tolerance. We deliberately do not claim usability gains without a user study.
References
- Observation uncertainty measures for sparse imputation
- Improving RNN Transducer with Target Speaker Extraction and Neural Uncertainty Estimation
- A deep generative model for gene expression profiles from single-cell RNA sequencing
- Uncertainty-aware single-cell annotation with a hierarchical reject option
- Xi-Vector Embedding for Speaker Recognition
- CellAgent: LLM-Driven Multi-Agent Framework for Natural Language-Based Single-Cell Analysis
- PIVOT: platform for interactive analysis and visualization of transcriptomics data
- TISSUE: uncertainty-calibrated prediction of single-cell spatial transcriptomics improves downstream analyses
- Combining packet loss compensation methods for robust distributed speech recognition
- Single-Cell Omics Arena: A Benchmark Study for Large Language Models on Cell Type Annotation Using Single-Cell Data