Automatic cell-type annotation tools label each cell in a single-cell RNA sequencing dataset, a measurement of which genes are active in each of thousands of individual cells. They typically fail in one of two ways. They force a confident-looking label onto a cell they do not understand, or they dump it into an "unassigned" bin that breaks every later analysis. We propose an annotator that returns three kinds of output instead. The first is a specific label when it is sure. The second is a coarser label, such as "T cell" rather than "CD4+ cytotoxic T cell", when only the coarser claim is safe. The third is a short list of review groups: clusters of deferred cells that a biologist inspects one group at a time.
What is new is how the thresholds get chosen. The user sets two numbers a bench scientist can actually reason about: the error rate they will tolerate among committed labels, and how many groups they are willing to review by hand. The system then picks its commit and deferral thresholds jointly, with a finite-sample guarantee on the first number, so that the reviews are spent where they fix the most labels. The claim to test is concrete. At a fixed number of human reviews, budget-aware joint calibration should yield more correct, more specific labels than the best existing recipe, which is hierarchical rejection followed by ad hoc sub-clustering of what was rejected.
Why this gap, and what already exists
Four lines of work show the ingredients exist but have never been assembled under a human-effort constraint.
The first line motivates the proposal and is its closest prior work. Uncertainty-aware single-cell annotation with a hierarchical reject option benchmarks forced, flat-reject and "partial" rejection on cell ontologies, where a cell is labeled with an ancestor when the leaf is uncertain. It finds that partial rejection keeps substantially more label information. Its thresholds are swept on accuracy-rejection curves rather than chosen with any guarantee, and rejected cells simply stop.
The second line supplies the guarantee. Hierarchical Selective Classification climbs a taxonomy until confidence clears a threshold chosen by split conformal prediction, a way of setting a cutoff on held-out labeled data so that an error bound holds with finite samples. It was tested only on ImageNet and iNaturalist, and it has no notion of what happens to the cells it declines to label.
The third line supplies the grouping idea. DC3 routes ambiguous images into an overclustering head, meaning it deliberately splits them into more clusters than there are classes. That is where its reported 2.4x speed-up in human re-annotation comes from. Its deferral threshold is a fixed 0.5 and its data are images.
The fourth line motivates the budget framing. The accessible LLM tools CellMaster and CellAgent show that humans in the loop help: CellMaster's advantage grows from 7.1% to 18.6% with feedback. But their decisions about when to ask are qualitative LLM judgments.
The decision-theoretic rough set view in Decision-Theoretic Rough Sets for Three-Way Decision-Making derives accept, reject and defer thresholds from a cost matrix. It constrains the design in a useful way, because the research file names the same weakness this proposal targets: nobody can objectively write down that matrix. We therefore use it as the lens rather than the interface. An (error rate, review budget) pair implicitly selects a point on the family of cost matrices, and we will report which costs a user's choice corresponds to.
Core Claim
Let each cell $x$ receive a predicted node $\hat y(x)$ in the cell ontology, obtained by climbing from the leaf while the node's aggregated probability is below a threshold $\theta$. Let $d(x)=1$ mean the cell is deferred to review because its highest-node confidence is below a second threshold $\tau$. On $n$ labeled calibration cells we choose $(\theta,\tau)$ so that
$$ \frac{1}{n+1}\Big(\sum_{i=1}^{n} \mathbf{1}[d(x_i)=0,\ \hat y(x_i)\not\supseteq y_i] + 1\Big)\le \epsilon . $$
Here $\hat y\not\supseteq y$ means the predicted node is not the true leaf or one of its ancestors, and $\epsilon$ is the user's tolerated error. The $+1$ correction is the standard conformal finite-sample term, and it gives the usual exchangeable-data guarantee.
Among the $(\theta,\tau)$ pairs that satisfy this bound, we pick the one that maximizes expected specificity after $B$ group reviews. Deferred cells are overclustered in expression space into review groups. The expected value of reviewing a group is estimated on calibration data as the number of cells whose label would become correct, or more specific, if the group received its majority label.
The bet is that the budget should change the thresholds. With a large $B$, it pays to defer more cells, because tight groups are cheap to fix. With $B=0$, the method reduces to conformal hierarchical selective classification. Groups formed in expression space also expose conflicts between the ontology and the transcriptome, a limitation named in the hierarchical-reject paper: such a conflict shows up as a review group whose cells span distant ontology branches. We report these groups as a diagnostic rather than silently forcing a label.
Research Program
We will build the annotator on top of the classical classifiers from the hierarchical-reject benchmark (logistic regression, linear SVM, random forest, with top-down and bottom-up aggregation). This makes any gain attributable to the decision layer rather than the backbone. We will also add a stronger neural encoder as a check that the effect survives a better base model.
The review simulator treats one review as revealing the majority true leaf of a group; every cell in the group then receives that label. This is a deliberately optimistic oracle. We will also run a noisy variant in which the revealed label is the majority ancestor at a random depth, so that results do not depend on a perfect expert.
The accessible layer is thin by design. It consists of two sliders (error tolerance, number of reviews), a ranked list of review groups with marker-gene dotplots, and a provenance record of the settings, in the spirit of PIVOT's data map. It can also serve as a tool that an agent such as CellAgent calls. In that role it replaces the Evaluator agent's qualitative accept/reject step with a quantitative stopping rule.
The theoretical caveat is stated upfront. Novel cell types absent from the calibration labels break exchangeability, so the error bound does not cover them. Our design routes low-confidence cells, including many novel ones, into review groups, and we measure empirically how often novel cells are nonetheless committed.
Baselines
The naive reference point is forced assignment of the top-probability leaf from the same classifier, together with flat rejection to "unknown" at the threshold that meets the same $\epsilon$. Together they show what the hierarchy and the guarantee each buy.
The strongest existing method is the partial-rejection pipeline from Uncertainty-aware single-cell annotation with a hierarchical reject option, run on its own datasets with its own classifiers. It is followed by Leiden sub-clustering of rejected cells into $B$ groups, which receive the same simulated reviews. This is the recipe a competent practitioner would assemble today. To be fair, it gets the same calibration split and the same classifier, and its threshold is tuned to the best final accuracy on a separate tuning split, so it enjoys a tuning advantage we do not. The file reports that this pipeline retains "substantially more label information than full rejection", but no reported result found on Lacuna for its exact retainment numbers, so we rerun it. For scale, Hierarchical Selective Classification reports a 14.94% hierarchical gain over standard selective classification on ImageNet-1k with its climbing rule; we reimplement that rule on cells as a second strong baseline.
For novel-type detection, mtANN is the strongest competitor, with its adaptive Gaussian-mixture threshold and unassigned bin. scPOT represents joint annotation and discovery. For neither does the file give numeric results (no reported result found on Lacuna for their AUPRC or accuracy values), so both are rerun on shared splits.
The key ablation is identical to our method except that $(\theta,\tau)$ are fixed first and the groups are formed afterward. This isolates the joint, budget-aware choice. A second ablation replaces overclustered review groups with random groups of equal size, to isolate the value of grouping.
Beating the strongest baseline means higher final hierarchical accuracy at matched specificity for every budget $B$ from 5 to 50 reviews. It also means the committed-label error staying at or below $\epsilon$ on held-out data, on at least four of the six benchmark datasets.
Why not just…?
Why not just run hierarchical selective classification with conformal thresholds on cells and stop there? This is the $B=0$ special case of our method, so it is a real threat. If review groups add little correct labeling per review over sending rejected cells to a biologist unsorted, the grouping contribution is empty. The random-grouping ablation and the curve of accuracy against $B$ settle this directly. DC3's 2.4x annotation speed-up on images is suggestive, but it is not evidence for transcriptomes.
Why not just reject, then sub-clustering the rejects, as practitioners already do? This is the most dangerous alternative, because it differs from our method only in whether the thresholds know the budget. scPOT argues that sequential annotate-then-cluster pipelines underperform because the two steps depend on each other, which motivates joint selection. If the fixed-threshold ablation matches the joint version within noise, we will conclude that the useful contribution is the guarantee plus the interface, and not the joint calibration.
Why not just use a mixture-model threshold like mtANN? mtANN chooses where to cut uncertainty scores without reference to what humans can afford to inspect, and it discards everything below the cut. We compare it at matched committed-error levels. We expect it to be competitive on detecting unseen types but to lose on how specific the final labels are once reviews are counted.
Why not just let an LLM agent decide when to ask the user, as CellMaster does? CellMaster's human-in-the-loop gains are real. Its confidence, however, is LLM-generated and carries no error bound. CellTypeAgent reports that adding domain text can lower LLM annotation performance, which makes self-reported confidence a weak basis for a stopping rule. Because API cost prevents a full rerun, we will compare on a subset of CellMaster's nine datasets. The metric is whether its "zoom-in" suggestions target the clusters our method ranks as highest value, and how their committed errors compare.
Proposed Experiments
The main experiment uses AMB, COVID, AzimuthPBMC, Flyhead, Flybody and PbmcBench from the hierarchical-reject benchmark. It plots final hierarchical accuracy and mean label depth against review budget $B$ at $\epsilon\in\{0.02,0.05,0.1\}$, averaged over five calibration splits. A second experiment uses mtANN's leave-one-cell-type-out pancreas and PBMC collections. It measures how often withheld types are committed as known labels versus concentrated into review groups, and how pure those groups are. A third experiment, on SEA-AD with 131 cell types, stress-tests a deep, fine-grained hierarchy and counts review groups that span distant ontology branches as detected conflicts. A small user study with bench biologists then compares time-to-final-labels using the two-slider interface against the partial-rejection output alone.
The decisive check is the joint-versus-sequential ablation. If budget-aware calibration does not beat fixed thresholds followed by sub-clustering at any budget, the core claim is false. The proposal would then shrink to a conformally guaranteed hierarchical annotator with a review interface. That is still useful, but it is a smaller result.
References
- Uncertainty-aware single-cell annotation with a hierarchical reject option
- Hierarchical Selective Classification
- A Data-Centric Approach for Improving Ambiguous Labels with Combined Semi-supervised Classification and Clustering
- CellMaster: Collaborative Cell Type Annotation in Single-Cell Analysis
- CellAgent: LLM-Driven Multi-Agent Framework for Natural Language-Based Single-Cell Analysis
- Decision-Theoretic Rough Sets for Three-Way Decision-Making in Dilemma Reasoning and Conflict Resolution
- PIVOT: platform for interactive analysis and visualization of transcriptomics data
- Cell-type annotation with accurate unseen cell-type identification using multiple references
- Realistic Cell Type Annotation and Discovery for Single-cell RNA-seq Data
- CellTypeAgent: Trustworthy cell type annotation with Large Language Models