acceptodds
Under review as a conference paper at ICLR 2027

scOLAR: Ontology-Anchored Open-Set Annotation of Single-Cell RNA-seq Data

Abstract

Reference-based single-cell annotation becomes an open-set problem when target data contain cell types absent from the reference. Existing approaches can reject unfamiliar cells, but the evidence used to distinguish known from unknown cells usually does not explicitly model cell-type hierarchy. This is limiting because an unseen fine-grained type may share broad transcriptional programs with a represented lineage, causing lineage similarity to support an incorrect known identity. Rejected cells must also be organized without knowing the number of unseen types. We present scOLAR, an ontology-anchored framework that separates fine-type evidence from broader-lineage affinity. scOLAR learns Cell Ontology-indexed prototypes with hierarchy-aware objectives, applies a reference-calibrated rule for selective label acceptance, groups rejected cells without a supplied novel-class count, and assigns broader ontology context post hoc. Across benchmarks, scOLAR achieves an AUROC of 0.973 for novelty detection, the highest published-protocol Overall accuracy on all five datasets, and the highest Novel accuracy on four. Under a separate common evaluator, scOLAR achieves a 7.2-percentage-point advantage in macro novel recovery together with higher known-cell accuracy. Matched ontology controls show stage-specific differences, with the largest observed changes in routing and post-rejection lineage context rather than novelty ranking or rejected-cell grouping. scOLAR turns open-set rejection into structured analysis of unresolved populations. Code is available in an anonymous repository.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single-cell RNA-seq, cell type annotation, open-set recognition, Cell Ontology, novel cell type discovery, prototype learning, hierarchical representation learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_TDpQOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper studies cell type annotation when the reference misses some target populations. scOLAR combines ontology-based prototype learning with a calibrated rejection rule, clusters rejected cells without a supplied class count, and adds broader lineage descriptions afterward. This separates fine label acceptance from interpretation of unresolved populations. Experiments across several single-cell datasets show strong novelty detection and accurate labels among retained known cells. Comparisons with scBOL show mixed advantages across recovery metrics. Ontology controls suggest that its effects are mainly on routing and lineage context rather than novelty ranking or clustering.

Critical strengths
  • Separating fine label acceptance, rejected-cell grouping, and lineage interpretation is sensible. In particular, adding a coarse description without silently converting a rejection into a known label addresses a real annotation problem.
  • The experiments distinguish the stages of the pipeline and report mixed results clearly. Figure 3 and the common-evaluator comparison are useful for understanding what the ontology changes and where the method still loses to scBOL.
Critical weaknesses
  • The motivation focuses on unseen types being confused with related known types, but the results do not show whether the hierarchy losses reduce that particular error. In Figure 3, real ontology has little effect on novelty ranking or grouping, and the star control remains competitive on correct-known routing. The specific benefit of learning biological hierarchy is still unclear.
  • Figure 5 illustrates a rejected astrocyte population whose type is already in the reference. This is a useful fallback after a false rejection. The benefit of this stage for genuinely unseen populations is still unclear.
Questions for authors
  1. Do the existing predictions show fewer false known assignments for unseen types that have close represented relatives? This would help connect Figure 3 to the main motivation.
  2. Can the existing LCC results separate genuinely unseen cells from falsely rejected known cells, and explain what useful lineage information is recovered for the former?
  3. Does the clustering algorithm affect final algorithm outcomes?
4 comments
AuthorsOpenReview ↗

Response to CW1/Q1 and CW2/Q2: Close-Relative Rejection and Genuine-Novel LCC

We are truly grateful for your careful review and thoughtful questions. We examined frozen predictions, ontology controls, ablations, and biological evidence for genuinely unseen populations.

CW1 / Q1. Close-relative false-known acceptance and hierarchy learning

HCL organizes source-prototype geometry, nSep reduces likely-novel affinity to mapped types, and LCC provides ancestry after rejection without changing the decision.

We define false-known acceptance (FKA) as the fraction of genuinely unseen cells with close represented relatives incorrectly accepted as mapped known. All ontology conditions use the same REAL-CL-defined stratum (shared-ancestor depth-gap ≤2), each with its own source-calibrated threshold. Four-seed rates are:

DatasetREAL-CLShuffled CLStar
Cao7.99%7.53%6.92%
Quake_10x0.149%0.167%0.188%
Quake_Smart-seq24.27%4.44%4.46%
Wagner1.44%1.67%1.57%
Zeisel_20180%0%0%

REAL lowers close-relative FKA against both controls on three datasets. Cao's reversal is concentrated in withheld visceral muscle (CL:0008007), which accounts for just 8.1% of close-novel cells but 82.4% of REAL's close FKA, mainly accepted as represented relative CL:0000187. Type-level FKA is 81.2% under REAL versus 76.6%/73.2% under shuffled/star; REAL sensory-neuron/neuron FKA is 0.20%/1.79%.

Independent geometry audits establish HCL's role. Full models achieve true-CL triplet ordering 0.996 versus 0.468 without HCL, and ontology–cosine Spearman 0.501 versus −0.027, with 20/20 paired improvements. Separately trained REAL/shuffled/star models achieve ordering 0.996/0.603/0.497 and near–far separation 0.140/0.019/0.001 on identical true-CL triplets; REAL leads 40/40 comparisons on both metrics. With nSep, close FKA is lower than without it in 19/20 paired runs (one tie), including Cao 7.99% versus 11.92% and Wagner 1.44% versus 3.52%. A post-hoc, target-label-assisted Known-retention-matched diagnostic retains 19/20 improvements; deployment thresholds remain source-calibrated.

These results directly address the close-relative false-known acceptance question. HCL learns biologically ordered prototype relationships, and nSep consistently reduces erroneous acceptance of unseen relatives. The Cao type-level audit identifies where remaining errors concentrate.

CW2 / Q2. Lineage context for genuinely unseen cells

We separated genuinely unseen N from falsely rejected represented K, using ground truth only after freezing routing, communities, and LCC. Four-seed rates below use all genuine-N cells as the denominator:

DatasetSource-reachableLCC assignedCompatible LCC
Cao58.41%26.12%23.81%
Quake_10x54.06%0%0%
Quake_Smart-seq2100%0%0%
Wagner14.75%30.53%0%
Zeisel_201888.05%0.24%0.048%

For Cao N, compatible coverage of 23.81% comprises 18.70% exact withheld-term and 5.10% strict-ancestor recovery; compatibility among assigned N is 91.04%. Assigned K attains 96.85% conditional compatibility. Among eligible rejected cells, LCC assigns 27.28% of N versus 55.20% of K. This asymmetry is consistent with selective source support. Represented K has stronger direct reference evidence, so LCC abstains more often on N while retaining high compatibility and allowing exact recovery when it assigns.

For withheld Cao sensory neurons, LCC returns neuron (CL:0000540) in 3/4 seeds (scores 0.422/0.424/0.408); the fourth abstains at 0.3999 under q=0.40q=0.40. Community Jaccard is 0.833–0.922. A literature-curated marker panel fixed before expression scoring favors neuronal over muscle/epidermal modules in 18/18 neuron-assigned Cao communities (16 novel-majority), versus 3/10 other-term-assigned and 21/66 unassigned communities. The markers corroborate the neuronal program; LCC adds a standardized reference-linked lineage term without forcing a fine identity.

At q=0q=0, Quake_10x reaches 16.69% compatible coverage, identifying selective gating as one contributor; Wagner selects incompatible candidates even without gating. Frozen ancestry replay reproduces 529/529 native outputs. Cao N compatible coverage is 23.81% under REAL versus 13.24% under shuffled ancestry (18.81% after assignment-mass matching).

We will clarify §1's motivation, summarize FKA, geometry, and component evidence in §4.4 while retaining original statistics, and strengthen §4.6/Figure 5 with genuine-N LCC, frozen replay, and Cao marker evidence, distinct from the existing astrocyte K fallback. Full paired data, post-hoc diagnostic details, marker sources and community scores, and null controls will appear in the Appendix. New and modified content will be highlighted in blue.

AuthorsOpenReview ↗

Response to Q3: Clustering Algorithm and Resolution Sensitivity

Thank you for this thoughtful question. Our controlled experiments show that clustering affects downstream PREDNOVEL grouping and LCC, while upstream open-set decisions remain invariant. We evaluated 20 frozen episodes (five datasets × four seeds), holding checkpoints, embeddings, MLS thresholds, rejected cells, and the original 15-NN graph fixed. All 20/20 upstream invariance audits passed. The same 594,685 target-cell–seed records were evaluated under each of six clustering settings. Novelty scores, Known/PREDNOVEL decisions, and accepted known labels were identical.

On the same graph, Louvain yielded comparable grouping performance to the original Leiden (resolution 1.0). Mean AMI was 0.6620→0.6643 (paired Louvain−Leiden Δ=+0.0023), and ARI was 0.3840→0.3949 (Δ=+0.0108). AMI improved in 12/20 paired episodes and ARI in 14/20. Mean community counts were 26.45 versus 25.15, with average per-episode median community sizes of 483.9 versus 494.7. Both algorithms grouped all rejected cells without noise or communities smaller than five cells.

The prespecified Leiden resolution sweep (0.5/0.8/1.0/1.2/1.5) revealed a systematic granularity effect. Across the extremes, mean community counts increased from 19.75 to 31.35, while AMI changed from 0.6964 to 0.6378 and ARI from 0.4696 to 0.3015. Relative to the original resolution 1.0, AMI increased at 0.5 in 20/20 episodes and decreased at 1.5 in 20/20. The original resolution was fixed without target-label tuning.

Clustering choice can also improve downstream lineage interpretation. On Cao genuinely novel cells, Louvain increased compatible LCC coverage from 23.81% to 27.40% (+3.60 pp) while maintaining high conditional compatibility among assigned N (91.04% versus 90.79%). Compatible coverage among falsely rejected K remained similar (53.42% versus 53.20%). N rates use all genuinely novel cells as the denominator, while K rates use rejected known cells. Effects varied across datasets. In Wagner, N assignment changed from 30.53% to 27.54%, but compatible N coverage remained zero under both algorithms. Both Quake datasets remained unassigned under the native LCC gate across the tested settings.

These results demonstrate a practical benefit of scOLAR's modular design. Clustering can adjust novel-group granularity and improve compatible lineage coverage, as observed on Cao, without retraining or changing Known/PREDNOVEL decisions. We plan to summarize the stage-specific findings alongside the grouping/LCC discussion, with complete paired AMI/ARI, community-size, resolution, and N/K-stratified LCC analyses in the Appendix. New and modified content will be highlighted in blue.

I appreciate authors' effort on the rebuttal. My major concerns have been resolved. However, I do think clutering is an important component in the method design as it acts as a role that controls novel-group granularity. As authors mentioned in the response, they plan to add these discussions into appendix. It would be interesting to examine what is the consequence of the granularity difference if combining LLMs with different clustering algorithms. As an update, I changed my soundness score to this manuscript. In summary, I think I am leaning to accept this paper.

Thank you so much for revisiting our responses and for your encouraging update! We're glad that our additional analyses have addressed your main concerns. We agree that clustering granularity deserves further attention, and your suggestion about combining LLMs with different clustering algorithms is an interesting direction for future work. We also appreciate your thoughtful feedback throughout the discussion!

The Authors

Reviewer_LvKhOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The authors introduce an ontology-anchored framework for open-set annotation of single-cell RNA-seq data under incomplete reference coverage, namely scOLAR. The method combines hierarchy-aware prototype learning with reference-calibrated selective classification, followed by class-count-free clustering of rejected cells and post-hoc lineage interpretation without altering the original predictions. The experiments demonstrate strong novelty-detection performance and provide a detailed analysis of selective-labeling trade-offs and how ontology structure influences different stages of the framework.

Critical strengths
  1. The problem is well motivated, and incomplete reference coverage is a practical challenge, especially for closely related unseen types. The framework has a clear modular design that separates selective annotation, class-count-free grouping, and post-hoc lineage interpretation.
  2. Novelty-detection performance is strong and annotation results are competitive across multiple datasets. I appreciate the effort to distinguish evaluation protocols and to report the mixed, statistically nonsignificant findings from the ontology controls.
Critical weaknesses
  1. The benefit of using the correct Cell Ontology structure is not fully clear. The matched ontology controls show modest, mixed effects, and none of the primary contrasts is statistically significant after correction. Novelty ranking and PREDNOVEL grouping remain largely stable across ontology conditions.
  2. The biological utility of LCC needs further clarification. It remains less clear how well LCC describes genuinely unseen cell types, and why assignment coverage is low or zero on some datasets.
Questions for authors
  1. What do the authors see as the main benefit of using the correct Cell Ontology structure, given the mixed results of the ontology controls?
  2. Could you briefly clarify the respective roles of HCL, DBR and nSep? Is there evidence already in the current experiments that helps distinguish them?
  3. Why is the full-head maximum logit used for rejection, while labels are assigned only from mapped source types?
  4. What motivates the fifth-percentile calibration rule? I wonder how sensitive known-cell retention and novel recall are to this choice.
  5. In what way does LCC complement conventional marker-gene analysis? What extra biological information do the ontology-based descriptions provide?

Minor Comments

  1. The method section is dense with abbreviations. A short overview of how the main objectives fit together would improve readability.
  2. The limitations paragraph in the main text is quite brief. A few of the qualifications already discussed in the appendix (especially the transductive setting) could usefully be brought into the main text.
4 comments
AuthorsOpenReview ↗

Response to CW1 and Q1–Q2: Cell Ontology and Component Contributions

Thanks for your thoughtful review and encouraging feedback on scOLAR's modular design and evaluation protocol! We conducted frozen ontology interventions, whole-ontology geometry audits, and 100-run component ablations to clarify the contributions of correct CL structure.

CW1 / Q1. Contribution of the Correct Cell Ontology Structure

Correct CL topology contributes most directly to hierarchy-consistent prototype geometry and source-supported lineage interpretation. Its effects on novelty ranking and PREDNOVEL grouping remain smaller and heterogeneous across datasets.

Frozen LCC replay isolates post-hoc ancestry. With 20 REAL-CL models, scores, routing, and communities fixed, we reproduced 529/529 native LCC outputs. Among genuinely novel Cao cells, native compatible lineage coverage was 23.81% under REAL, versus 13.24% under depth-shuffled ancestry and 18.81% after approximately matching assignment mass (+10.57/+5.00 pp). Matching used all rejected cells without access to novel ground-truth labels.

In a separate 30-permutation ontology-label null, we shuffled labels within ontology-depth strata and transported the full ancestor closure, keeping source identities, logits, and native-assigned communities fixed. This diagnostic used identical communities with forced candidate selection, without reapplying the native LCC gate. REAL yielded 23.73% compatible coverage versus a 0.29% null mean (mean difference +23.45 pp, range +15.12 to +23.73 pp; 30/30 positive). This fixed-set null and the native gated replay answer distinct questions.

Independent topology-geometry audit. Separately trained REAL/depth-shuffled/star models were evaluated on identical true-CL triplets across five datasets × four seeds:

Training ontologyTriplet ordering ↑Spearman ↑Near–far separation ↑
REAL CL0.9960.5010.140
Depth-shuffled CL0.6030.1180.019
Generic-star CL0.4970.0460.001

REAL led in 40/40 paired comparisons for both ordering and near–far separation, and 36/40 for Spearman. Separately trained ontology controls improved joint correct-known routing on Cao (+4.35 pp) and Zeisel (+3.95 pp) versus shuffled CL, while Wagner decreased by 1.79 pp. The original six prespecified ontology-control contrasts retain their reported nonsignificant multiplicity-adjusted results. These new audits directly isolate geometry and lineage compatibility, which the primary prediction/grouping endpoints do not.

Q2. Roles of HCL, DBR, nSep, and Reconstruction

The four objectives optimize distinct quantities, with potentially overlapping downstream effects through a shared encoder. We evaluated 100 models (five datasets × four seeds × Full and four one-loss-out conditions).

ObjectiveDirect optimization roleTargeted evidence
HCLCL-defined source-prototype triplet orderingRemoving HCL changed ordering 0.996→0.468 and ontology–cosine Spearman 0.501→−0.027 (20/20 paired improvements for Full).
DBRFine-versus-coarse source-cell affinity marginFull raised fine-prototype cosine by 0.0081 and multitype coarse-centroid coherence by 0.0067 versus w/o DBR (20/20 for both).
nSepPenalizes likely-novel affinity to mapped source typesRemoving nSep lowered AUROC by 0.0183 (20/20) and increased close-relative false-known acceptance in 19/20 pairs (one tie): Cao 7.99%→11.92%; Wagner 1.44%→3.52%.
ReconstructionPreserves expression information for representation and groupingRemoving it lowered PREDNOVEL AMI by 0.0596 across all five datasets; close-relative false-known acceptance also rose in 20/20 pairs.

DBR's mean fine–coarse gap changed by only −0.0012. Singleton coarse groups covered 19.9%–77.4% of source cells and explained 84.7%–100% of the observed hinge magnitude: their fixed positive margin is structurally unattainable. HCL and DBR depend on CL relationships; nSep and Reconstruction do not directly use ontology topology. In aggregate ablations (Δ=Ablated−Full\Delta=\mathrm{Ablated}-\mathrm{Full}), removing HCL changed AUROC/AMI by +0.00149/+0.00041; removing DBR by +0.00040/+0.00202. Their direct geometry/affinity effects are more consistent than their aggregate prediction effects.

HCL shapes the source prototypes used by known-cell routing; the independent REAL-CL routing differences above show where this geometry is used downstream. The frozen LCC intervention separately identifies the effect of correct ancestry at the interpretation stage.

We will incorporate the component and topology audits into §4.4 while retaining the original Holm-adjusted statistical conclusions.

AuthorsOpenReview ↗

Response to CW2 and Q3–Q5: MLS Calibration and Biological Utility of LCC

We also examined MLS scoring, calibration sensitivity, and the biological utility of LCC.

Q3. Full-Head MLS and Mapped-Source Label Selection

Acceptance uses the maximum over all ontology-indexed prototypes, Mifull=max⁡j∈Oℓi,jM_i^{\mathrm{full}}=\max_{j\in\mathcal O}\ell_{i,j}; the returned label maximizes ℓi,k\ell_{i,k} over mapped source types k∈Kk\in\mathcal K only. Full-head scoring follows the trained classifier and reference-only calibration. Unobserved rows compete in source cross-entropy without positive source labels; mapped-only selection avoids unsupported zero-shot assignments.

At a fixed threshold, Mifull≥MimappedM_i^{\mathrm{full}}\ge M_i^{\mathrm{mapped}} makes full-only acceptance possible; a mapped label could then be returned below its acceptance threshold. Across 20 frozen REAL-CL runs, 0/594,685 target cell–run evaluations exhibited full-only acceptance. In Zeisel, 163,422 evaluations had different full/mapped scores, yet all were rejected under both rules. Independently recalibrating each head at five reference quantiles yielded 100/100 identical thresholds and Known/PREDNOVEL decisions. The theoretical mismatch did not occur at the tested thresholds.

Q4. Fifth-Percentile Calibration and Sensitivity

We perform 20 source-class subsamples, compute each reference MLS quantile, and use the median threshold, without target labels. The following sensitivity results come from the 20 REAL-CL ontology-control runs:

Reference quantileKnown retentionNovel recall
1%83.59%96.82%
2.5%78.83%97.41%
5%74.18%97.80%
10%69.40%98.07%
20%60.76%98.42%

The fixed reference-only 5% point retained 74.18% known cells with 97.80% novel recall. The 1%–20% sweep shows a gradual trade-off without a sharp optimum at 5% or target-label tuning. This MLS quantile differs from the LCC gate (q=0.40q=0.40).

CW2 / Q5. Genuine-Novel LCC and Complementarity to Marker Analysis

LCC supplies a source-supported, CL-indexed lineage term to unresolved communities without changing their PREDNOVEL status. Ground-truth labels were used only after native outputs were frozen to separate genuinely unseen N from falsely rejected known K. Four-seed rates below use all N cells as denominator:

DatasetSource-reachableAssignedCompatible
Cao58.41%26.12%23.81%
Quake_10x54.06%0%0%
Quake_Smart-seq2100%0%0%
Wagner14.75%30.53%0%
Zeisel_201888.05%0.24%0.048%

For Cao genuinely novel cells (N), 23.81% received compatible lineage terms: 18.70% exact withheld-term and 5.10% strict-ancestor recovery (four-seed means). Conditional compatibility among assigned N was 91.04%. Among eligible rejected cells, assignment covered 27.28% of N versus 55.20% of falsely rejected known cells (K); assigned K achieved 96.85% conditional compatibility.

In a withheld sensory-neuron example, the subset-specific ciliated-neuron marker osm-6 appeared in only 5/639 cells (0.78%); this marker alone gives limited coverage of the fine sensory identity. Pan-neuronal expression supported a broader neuronal program, and LCC assigned neuron (CL:0000540) in 3/4 seeds (scores 0.422/0.424/0.408); the fourth abstained at 0.3999 under q=0.40q=0.40.

Across 94 Cao seed–communities, 18 received neuron (16 novel-majority). A prespecified literature-curated panel found stronger neuronal than muscle/epidermal modules in 18/18 neuron-assigned, versus 3/10 other-term-assigned and 21/66 unassigned communities. Markers corroborate the program; LCC additionally supplies a standardized reference-linked lineage term for ontology-based analysis without forcing a fine identity when evidence is insufficient.

Coverage reflects three separate requirements: source-reachable ancestry, compatible candidate selection, and selective gating. Quake datasets have reachable ancestry but no assignments at q=0.40q=0.40; lowering the gate to q=0q=0 yields 16.69% compatible coverage on Quake_10x and 32.45% on Zeisel (versus 0.048% native). Wagner assigns 30.53% of N but selects incompatible candidates even at q=0q=0. In Cao, gating raises conditional compatibility from 46.73% (q=0q=0) to 91.04%, illustrating the value of selective abstention.

Minor comments and manuscript revisions. We will add a concise objective-to-inference overview in §3.2 and clarify MLS scoring and calibration in §3.3. Section 4.4 will summarize the component ablations, ontology geometry, and close-relative diagnostics while retaining the original statistical conclusions. Section 4.6 and Figure 5 will present genuine-novel LCC, frozen ancestry controls, and Cao marker evidence. We will expand §5 to clarify the transductive setting and selective LCC coverage, with full experimental details in the Appendix. The changes will be highlighted in blue.

The author's detailed response addressed my concern. I am very happy to recommend acceptance. Good luck!

We are pleased to know that our responses have addressed your concerns. Please do not hesitate to reach out to us if you have any further questions.

the authors

Reviewer_QT3POpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies open-set cell-type annotation in single-cell RNA sequencing data, where the target dataset may contain cell types absent from the reference. The authors propose scOLAR, an ontology-guided framework that incorporates hierarchical cell-type relationships into reference-based annotation and post-hoc lineage interpretation of rejected cells. The proposed approach uses hierarchy-aware representation learning to detect novel cells, clusters rejected cells, and assigns ontology-based lineage information to the resulting groups. Experiments across five benchmark datasets show improvements over existing methods in distinguishing known and novel cell types, maintaining annotation accuracy, and recovering previously unseen cell populations. Additional analyses show the contributions of ontology information to routing and the biological interpretation of rejected cells.

Critical strengths

Incorporating the hierarchical structure of the Cell Ontology into open-set annotation is a well-motivated approach that addresses a practical problem. The experimental evaluation is extensive, covering a range of datasets and settings. The results demonstrate that scOLAR achieves competitive or superior performance compared to a broad set of existing methods.

Critical weaknesses

The contribution of the Cell Ontology to the overall performance is unclear. In Table 15, the real ontology provides no consistent improvement over the shuffled or star controls in novelty ranking, known-cell annotation, or grouping of rejected cells. Although the authors suggest that ontology information benefits post-hoc lineage interpretation, the LCC results in Table 14 show limited coverage and inconsistent compatibility across datasets. These results raise questions about how much the actual ontology structure contributes to the proposed framework. A component-wise ablation that removes each loss one at a time is needed to clarify their individual contributions.

The acceptance and label selection rules in Eq (7) are not clearly explained. The Known/PredNovel decision uses the maximum logit over all ontology-indexed prototypes, but an accepted label is selected only from the mapped source types. Therefore, a cell could be accepted due to a high score on an unobserved prototype, which receives no positive source supervision, even when its scores for all mapped source types are below the threshold. For a fixed model and threshold, including unobserved prototypes can only increase the number of accepted cells. A comparison between full-head MLS and MLS over mapped source types would help clarify this issue.

Questions for authors

See weaknesses

3 comments
AuthorsOpenReview ↗

Response to Critical Weakness 1: Cell Ontology and component contributions

We thank the reviewer for their careful review and constructive feedback. We investigated ontology geometry, genuine-novel lineage interpretation, and the contributions of individual losses.

C1. Contribution of correct CL structure

HCL shapes ontology-consistent geometry. Across five datasets × four seeds, Full versus w/o HCL achieved triplet ordering 0.996 vs 0.468 and ontology–cosine Spearman 0.501 vs −0.027 (20/20 paired improvements). Full increased near-prototype cosine by 0.131 and decreased far-prototype cosine by 0.013; removing DBR largely preserved ordering.

Separately trained REAL/depth-shuffled/star models, evaluated on identical true-CL triplets, achieved ordering 0.996/0.603/0.497 and near–far separation 0.140/0.019/0.001. REAL led in all 40 paired comparisons for both ordering and near–far separation (Spearman: 36/40).

Routing effects vary: REAL minus shuffled improved joint correct-known on Cao (+4.35 pp) and Zeisel (+3.95 pp) but decreased it on Wagner (−1.79 pp). All six original primary comparisons remain nonsignificant after correction (Table 15). Geometry and downstream metrics reflect different pipeline stages; C2–C3 further examine these stage-specific effects.

C2. LCC on genuinely unseen types

We separated genuinely novel from falsely rejected known cells, retaining native LCC (q=0.40q=0.40). Coverage uses all novel cells; reachability is an offline structural diagnostic.

DatasetReachableAssignedCompatible
Cao58.41%26.12%23.81%
Quake_10x54.06%0%0%
Quake_Smart-seq2100%0%0%
Wagner14.75%30.53%0%
Zeisel_201888.05%0.24%0.048%

Cao demonstrates genuine-novel lineage recovery: 91.04% compatibility among assigned novel cells and 23.81% compatible lineage coverage across all genuinely novel cells. Among 94 eligible seed–communities, 28 received terms. Of 18 neuron-assigned communities, 16 were novel-majority. A withheld sensory-neuron community received the compatible ancestor neuron (CL:0000540) in 3/4 seeds (scores 0.422/0.424/0.408); the fourth abstained at 0.3999. PREDNOVEL routing was unchanged. Literature-curated neuronal modules dominated muscle/epidermal modules in 18/18 neuron-assigned communities versus 3/10 other-term-assigned and 21/66 unassigned; the panel was fixed before scoring.

Frozen ancestry replay isolates interpretation. Fixing 20 REAL models, routing, scores, and Leiden communities reproduced 529/529 native outputs. Cao genuine-novel compatible coverage was 23.81% REAL, 13.24% shuffled, and 18.81% shuffled with approximately matched assignment mass (REAL: +10.57/+5.00 pp). Matching used all rejected cells without ground-truth novel labels. Correct ancestry improves descriptions with predictions fixed.

Failure-mode analysis explains cross-dataset heterogeneity. Quake datasets have reachable ancestry but no communities pass the q=0.40q=0.40 gate; at q=0q=0, Quake_10x reaches 16.69% compatible coverage and Zeisel rises from 0.048% to 32.45%. Wagner assigns 30.53% of novel cells but selects incompatible candidates even at q=0q=0. In Cao, the gate raises conditional compatibility from 46.73% (q=0q=0) to 91.04%. Native results retain q=0.40q=0.40.

C3. Contributions of the four objectives

We completed 100 runs (five datasets × four seeds × Full and four one-loss-out models); Δ=Ablated−Full\Delta=\mathrm{Ablated}-\mathrm{Full}.

ObjectiveDirect rolePaired evidence
HCLSource-prototype hierarchyOrdering 0.996 vs 0.468 (20/20); Δ\DeltaAUROC +0.00149, Δ\DeltaAMI +0.00041
DBRFine/coarse source-cell affinityFine cosine +0.0081; multitype coarse coherence +0.0067 (both 20/20); Δ\DeltaAUROC +0.00040, Δ\DeltaAMI +0.00202
nSepLikely-novel/source separationΔ\DeltaAUROC −0.0183 (20/20); close-relative false-known acceptance increased (19/20, one tie)
ReconstructionExpression preservation and groupingΔ\DeltaPREDNOVEL AMI −0.0596 (all five datasets)

DBR changes the mean fine–coarse gap by only −0.0012; singleton coarse groups account for 84.7%–100% of observed hinge magnitude. HCL and DBR exhibit consistent geometry- and affinity-level effects, respectively, despite small, heterogeneous changes in aggregate AUROC and AMI. The objectives share an encoder, so effects overlap: removing reconstruction also increased close-relative false-known acceptance in 20/20 pairs. HCL shapes geometry; nSep directly promotes rejection; reconstruction supports grouping, without attributing nSep's benefit to CL topology.

We will add compact component/geometry evidence to §4.4, retaining original statistics, and genuine-novel LCC, frozen replay, and Cao marker evidence to §4.6/Figure 5. Full paired results, null controls, and diagnostics will appear in the Appendix. New and modified content will be highlighted in blue.

AuthorsOpenReview ↗

Response to Critical Weakness 2: Full-Head MLS and Label Selection

We appreciate the reviewer's precise observation about Eq. (7). We directly tested whether a high logit from an unobserved prototype could cause acceptance without sufficient mapped-source evidence.

C1. Full-head confidence and mapped-source labels

Equation (7) separates the rejection statistic from the accepted label. For ontology-indexed prototypes O\mathcal O and mapped source types K⊂O\mathcal K\subset\mathcal O,

Mifull=max⁡j∈Oℓi,j,Mimapped=max⁡k∈Kℓi,k.M_i^{\mathrm{full}}=\max_{j\in\mathcal O}\ell_{i,j},\qquad M_i^{\mathrm{mapped}}=\max_{k\in\mathcal K}\ell_{i,k}.

We accept when Mifull≥αM_i^{\mathrm{full}}\ge\alpha and then return arg⁡max⁡k∈Kℓi,k\arg\max_{k\in\mathcal K}\ell_{i,k}; otherwise the cell enters PREDNOVEL. The full-head score follows the ontology-indexed classifier used for source training and reference-only calibration. Unobserved prototypes participate in the full classifier and cross-entropy normalization but receive no positive source labels. Restricting the returned identity to K\mathcal K prevents unsupported zero-shot fine-type labels; rejected cells can receive broader context later through LCC.

The reviewer correctly identifies a mathematical possibility: Mifull≥MimappedM_i^{\mathrm{full}}\ge M_i^{\mathrm{mapped}}, so, at a fixed threshold, full-head acceptance can only be a superset of mapped-head acceptance. In a full-only acceptance, Eq. (7) would indeed return the best mapped label even though its logit would be below the acceptance threshold. We tested precisely this failure condition.

C2. Fixed-threshold cell-level audit

Following the reviewer's suggestion, we compared full-head MLS with MLS over mapped source types in 20 frozen REAL-CL runs (five datasets × four seeds) with logits and the original reference-calibrated thresholds fixed.

ComparisonEvaluationsOutcome
Full-only Known acceptance594,685 target cell–run evaluations0 cases
Distinct full/mapped scores on Zeisel163,422 evaluationsAll rejected under both rules

Thus, no evaluated cell was accepted solely because an unobserved prototype had the highest logit. The rule of three gives a descriptive 95% upper bound of 3/594,685≈5.0×10−63/594{,}685\approx5.0\times10^{-6} per evaluation; cell–run evaluations are not independent biological replicates. The Zeisel result confirms that the heads can return different scores, but these differences did not change any acceptance decision at the deployed threshold. This pattern is consistent with unobserved rows competing in source cross-entropy without receiving positive source supervision: their logits can lead among rejected target cells yet remain below the calibrated threshold.

C3. Independently recalibrated head comparison

We then calibrated each head separately using the same 20 reference-only source-class subsamples per run and median aggregation, without target labels. At reference quantiles 1%, 2.5%, 5%, 10%, and 20%, all 100/100 dataset–seed–quantile comparisons yielded identical thresholds and Known/PREDNOVEL decisions. Across this range, 20-run mean known-cell retention changed from 83.59% to 60.76% and novel recall from 96.82% to 98.42%, while full-head and mapped-head decisions remained equal at every tested operating point.

Together, the fixed-threshold and recalibrated audits show that the reviewer's mathematically valid full-only acceptance scenario did not occur in any evaluated dataset at the tested thresholds. The mapped-source constraint also ensures that every accepted identity remains a reference-supported type.

We will clarify the full-head confidence score and mapped-source label selection following Eq. (7) in §3.3. Appendix C.1 will include the fixed-threshold and recalibrated head-choice audits, together with the reference-quantile sensitivity analysis. These additions will be highlighted in blue.

Dear Reviewer QT3P,

Thank you again for your insightful feedback. We have taken your concerns seriously and conducted additional experiments and analyses to address them. Guided by your suggestions, these results help clarify the specific contributions of the Cell Ontology and directly resolve the potential inconsistency in the acceptance and label-selection rules.

We are incorporating these findings and clarifications into the revised manuscript. We truly appreciate the time and effort you have devoted to reviewing our work, and would be grateful to hear whether our responses have adequately addressed your concerns or if any points require further clarification.

Thank you for your consideration.

Best regards, The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.