acceptodds
Under review as a conference paper at ICLR 2027

SPORE: Semantic-Aware Global Post-Training for Reusable Single-Cell Representations

Abstract

Pretrained single-cell foundation models encode broad biological variation, yet their representations are not explicitly organized to make cell identity, biological context, and gene semantics accessible for downstream use. We introduce SPORE, a semantic-aware global post-training framework that adapts a pretrained single-cell foundation model once with heterogeneous biological supervision and yields a reusable checkpoint for deployment across studies and tasks. SPORE couples gene-semantic embedding alignment, ontology-aware cell supervision, and structured biological-context alignment to expose complementary biological semantics across scales. We evaluate the resulting checkpoint without target-specific backbone fine-tuning across complementary tests of downstream transfer, representation structure, and direct semantic inference. Across controlled evaluations, SPORE improves downstream transfer, makes biological information more accessible in frozen representations, and supports direct cell-type inference in new studies. These results support semantic-aware post-training as a means of organizing complementary biological semantics into a reusable single-cell representation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single-cell transcriptomics, representation learning, post-training, foundation models

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_jPTPOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3(+1)
Presentation
2
Soundness
2
Read review
Paper summary

SPORE proposes post-training a single-cell foundation model once to produce a checkpoint that can be reused across studies and tasks without further backbone fine-tuning. It combines gene-semantic alignment using GenePT, ontology-aware cell-type supervision, and biological-context alignment using biomedical text embeddings. Its main contribution is integrating these complementary signals into a shared representation that supports both lightweight downstream predictors and direct cell annotation.

The authors evaluate five backbones across six downstream tasks, including cell-type annotation, cancer classification, drug sensitivity, gene-level prediction, and perturbation prediction. They also examine representation geometry, performance with limited downstream labels, and direct annotation on 11 study-disjoint datasets. Supervision ablations and data-scaling experiments help assess where the improvements come from.

The clearest gains are in cell-type transfer and cancer classification, with additional evidence of improved label efficiency and closer alignment with ontology structure. Gene-level and perturbation improvements are smaller or mixed, and the combined objective does not consistently outperform individual supervision branches. Overall, the results support the usefulness of global semantic post-training, particularly for cell-level applications.

Critical strengths

The paper addresses a practical limitation of single-cell foundation models: adapting a model separately for every downstream setting limits its reuse. SPORE combines biological supervision at the gene, cell, and context levels in a single post-training stage, then reuses the resulting checkpoint without target-specific backbone updates. This makes the contribution relevant beyond a single annotation task and offers a practical way to improve existing pretrained models.

The evaluation covers five backbones and several downstream tasks, with consistent improvements in cell-type hAcc and cancer-classification AUROC across the paired backbones. The study-disjoint annotation experiments and label-efficiency analyses provide complementary evidence that the benefits extend beyond fitting the post-training labels. This breadth strengthens confidence in the approach, while the mixed gene-level and perturbation results help clarify where its benefits are strongest.

Critical weaknesses

The evidence for the benefit of combining all three supervision branches is limited. In Table 3, ontology-only supervision nearly matches the full objective on cell-type transfer and cancer classification, while context-only supervision performs better on drug sensitivity, and several individual branches perform better on tissue specificity. But I believe that shouldn't be the reason to reject this paper simply and bluntly. On the contrary, this leads me to believe that these improvements and results are highly credible. It is worth noting that many current single-cell foundation models employ evaluation methods that favor them—sometimes even inflating their reported performance. This kind of honest reporting is precisely what the field needs.

Questions for authors

That said, there are still many critical issues that should be solved or explained further:

  1. The post-training objectives may suppress biological variation not captured by the supplied annotations. In particular, cell-type supervision may encourage stronger separation between annotated populations, potentially obscuring the continuous transitions and developmental relationships within differentiating lineages. These relationships are not fully represented by the cell ontology hierarchy. Consequently, improved classification performance could come at the expense of preserving the connectivity of the underlying biological manifold. The current evaluations do not directly address this possibility. An additional analysis of trajectory inference or differentiation would help assess whether post-training preserves continuous developmental structure.

  2. Using gene-semantic alignment as a post-training task should be approached with caution. If GenePT embeddings are weak on the relevant tasks, aligning pretrained gene representations to them may constrain rather than improve those representations. Their informativeness also depends on the underlying NCBI descriptions (used in GenePT), which can be sparse for poorly characterized genes. For example, the summary for the protein-coding gene C16orf78 states only “Located in nucleus.” Such descriptions provide limited gene-specific functional information. These concerns suggest that GenePT embeddings may be more suitable for initializing the gene dictionary than serving as fixed post-training targets, although a direct comparison is needed to establish this.

  3. In the cell type annotation task across different experiments, it reports hAcc as the evaluation metric. However, as SPORE incorporates a hierarchy-related post-training task, ontology-aware cell identity, the improvement may be caused by the alignment between training task and evaluation metrics. So, it will be more persuasive if the authors can replace it as standard macro-F1 or accuracy.

  4. The color coding in panels A and B of Figure 4 is very difficult to interpret on first inspection. The legend placement makes it unclear which legend applies to each panel and whether colors have the same meaning across panels. Placing the relevant legend beside each panel, or explicitly identifying the panels associated with each legend, would improve readability.

3 comments
AuthorsOpenReview ↗

Official Response to Reviewer jPTP

Thank you for recognizing the practical value of a reusable checkpoint and the importance of transparently reporting task-dependent results. Our response is organized into three parts. Part 1 examines the task-dependent contributions of joint supervision and whether post-training preserves continuous developmental structure. Part 2 evaluates the quality and utility of GenePT embeddings as gene-semantic supervision targets. Part 3 addresses hierarchy-compatible versus exact-match evaluation and the presentation of Figure 4.

Part 1 — Joint supervision and developmental structure

W1 — Combining the supervision branches.

We appreciate the reviewer's thoughtful assessment of our mixed results and emphasis on transparent evaluation. We agree that ontology-only supervision accounts for much of the cell-type and cancer gains, while individual branches can outperform Full on specific tasks (Table 3). These findings motivate a closer examination of the relationships between supervision objectives and downstream tasks.

Task-dependent contributions. We added Geneformer leave-one-out experiments (Appendix E.4.1, Table A19; Figure A6A). Removing Cell reduces cell-type hAcc from 0.822 to 0.771 and Cancer AUROC from 0.743 to 0.721, while removing Context or Gene reduces tissue accuracy from 0.814 to 0.797 or 0.783, respectively. Interestingly, although Context-only achieves higher tissue accuracy than Full (0.881 vs. 0.814), removing Context from Full still reduces performance. This illustrates that an objective's standalone performance does not necessarily predict its contribution within joint training, revealing non-additive, task-dependent effects of combining supervision sources. Complementary gradient diagnostics (Figure A6B) further characterize their optimization interactions, showing weak positive Cell–Context alignment and nearly orthogonal Gene-related gradients on average.

Cross-study generalization. SPORE aims to consolidate heterogeneous biological supervision into one checkpoint reusable across studies and tasks. On 10,500 cells from 11 held-out studies, Full improves direct-annotation macro-F1 over ontology-only post-training from 0.3343 to 0.3658 (paired study-bootstrap Δ\Delta +0.0315, 95% CI [0.0008, 0.0492]; Appendix E.3, Table A18). Together, these results provide evidence of additional cross-study benefits while revealing meaningful task-dependent trade-offs in joint semantic post-training.

Q1 — Continuous developmental structure.

We thank the reviewer for raising the preservation of continuous developmental structure.

We investigated the preservation of continuous developmental structure through two trajectory-conservation evaluations comparing the original Frozen and SPORE-post-trained Geneformer checkpoints. SPORE uses an ontology-aware classification objective to make cell identity accessible from the learned representations, without explicitly constraining their pairwise geometry. This leaves room for the representations to retain variation beyond the supervised cell-type labels.

Trajectory conservation is measured by the scIB score (ρ+1)/2(\rho+1)/2 (higher is better), where ρ\rho is the Spearman correlation between reference and embedding-derived pseudotime:

Oetjen erythroid differentiation (Oetjen et al., 2018; Luecken et al., 2022). Using 2,629 cells and the scIB benchmark's fixed continuous DPT reference, we obtained:

EvaluationFrozenSPORE
Three-sample macro0.72120.7668
All cells pooled0.69430.7583
Sample A0.81510.8401
Sample P0.57010.6513
Sample U0.77840.8091

Trajectory conservation improves in all three sample-batches.

Ranzoni Smart-Seq2 fetal hematopoiesis (Ranzoni et al., 2021). As an additional cross-platform test, we evaluated 2,070 cells against the original authors' saved continuous DPT reference:

EvaluationFrozenSPORE
11-sample macro0.56330.5690
All cells pooled0.69110.7118

Both checkpoints retain all cells in a single connected component with finite inferred pseudotime in each cohort. Across these two cohorts, SPORE maintains or improves agreement with the reference developmental trajectories, with no observed loss of global graph connectivity. Complete evaluation protocols, sample-level results, and trajectory visualizations are provided in the revised appendix.

The added trajectory protocols, results and visualizations are in Appendices F.3–F.4, Tables A25 and A27 and Figures A9–A10.

AuthorsOpenReview ↗

# Part 2 — Gene-semantic target quality

Q2 — Gene-semantic target quality.

We appreciate the reviewer’s caution regarding the quality of GenePT as a fixed supervision target.

Target quality. To our knowledge, GenePT is among the most comprehensive text-derived gene-level knowledge bases available, providing broad coverage of NCBI descriptions. Our audit found that 88.65% of covered Geneformer genes have archived descriptions exceeding ten words, while 11.35% have short or unavailable descriptions, including C16ORF78. Looking ahead, the strong results on protein-informed backbones (Table 1) motivate exploring protein-derived knowledge as another promising source of semantic priors.

Functional evidence. We further compared Frozen Geneformer, raw GenePT, and Gene-only post-training under coverage-matched evaluation protocols (289 common genes for Gene–Tissue; coverage-matched edges for UGRN):

TaskFrozenRaw GenePTGene-only
Gene–Tissue accuracy ↑\uparrow0.82410.60000.8517
UGRN AUPRC ↑\uparrow0.67910.72140.6858

Raw GenePT exhibits task-dependent utility, weaker on tissue but stronger on UGRN (including all STRING endpoints). Notably, Gene-only alignment improves tissue accuracy (+0.0276) and UGRN AUPRC (+0.0068) over Frozen. Thus, an external semantic prior need not be a superior standalone representation to provide useful auxiliary supervision.

Design rationale. SPORE aligns a learned projection of pretrained gene embeddings with GenePT, allowing the native representations to retain information beyond the external prior. Supervision also requires only partial gene coverage, permitting future quality-aware target selection.

Regarding the alternative of GenePT-based initialization, we agree that this is a plausible strategy, particularly when constructing a backbone from scratch, where the gene dictionary and backbone can be optimized jointly.

In the post-training of an existing backbone, however, GenePT-based initialization presents two related challenges. First, GenePT embeddings (1,536 dimensions) and the backbone's native gene embeddings (256 dimensions in Geneformer) differ in dimensionality, requiring an additional mapping step. More fundamentally, matching their dimensions does not ensure compatibility with the representation space expected by the pretrained backbone.

To illustrate this compatibility challenge, we tested a ridge-based GenePT embedding-replacement control on Geneformer. We mapped normalized GenePT embeddings to the corresponding original Geneformer gene embeddings for the 20,824 covered genes and replaced the original entries with the mapped vectors. The mapping achieved a mean cosine similarity of 0.6338 (median 0.6488) and a relative L2 reconstruction error of 0.7640, indicating substantial differences from the original gene representations.

We further compared the replacement variant with the original-vocabulary Cell/Context-only control (−Gene) under the same training budget (seed 37, 6,546 steps), using the native annotation head on the same 15,190 donor-held-out test cells:

Metric−GeneGenePT replacement
Accuracy ↑\uparrow0.63500.5520
Macro-F1 ↑\uparrow0.31590.2347
hAcc ↑\uparrow0.81410.7481
hMacro-F1 ↑\uparrow0.46150.3595

The replacement variant underperforms the original-vocabulary control across all four metrics, illustrating the practical challenges of directly incorporating external gene embeddings into an existing pretrained backbone. Developing more compatible ways to integrate complementary gene-semantic information while preserving pretrained representations remains an interesting direction for future work.

AuthorsOpenReview ↗

Part 3 — Evaluation metrics and figure readability

Q3 — Hierarchy-compatible versus conventional evaluation.

We thank the reviewer for raising this important concern. We use hAcc and hMacro-F1 to accommodate differences in annotation granularity, following scTab and HCE-scTab, where valid descendant predictions are not penalized for being more specific than reference annotations. To assess potential metric alignment, we additionally report standard exact-match accuracy and macro-F1 on the original frozen-feature transfer benchmark (revised Table A5):

BackbonehAcc (Frozen →\rightarrow SPORE)Accuracy (Frozen →\rightarrow SPORE)Macro-F1 (Frozen →\rightarrow SPORE)
Geneformer0.649 →\rightarrow 0.8220.546 →\rightarrow 0.6470.167 →\rightarrow 0.323
scCello0.747 →\rightarrow 0.8330.683 →\rightarrow 0.7290.269 →\rightarrow 0.321
TEDDY-G0.665 →\rightarrow 0.8240.572 →\rightarrow 0.6740.262 →\rightarrow 0.404
UCE0.839 →\rightarrow 0.8740.781 →\rightarrow 0.8450.462 →\rightarrow 0.564
TF-Sapiens0.871 →\rightarrow 0.8940.841 →\rightarrow 0.8020.579 →\rightarrow 0.492

Both conventional metrics improve on four of five backbones. For TF-Sapiens, the proportion of valid but more-specific-than-reference predictions increases from 3.0% to 9.2%. This may reflect a backbone-specific interaction between autoregressive pretraining and ontology-aware supervision that favors finer-grained predictions.

We further evaluated frozen Geneformer representations on two independent datasets with more consistent, flat cell-type annotations:

DatasetMacro-F1 (Frozen →\rightarrow SPORE)Accuracy (Frozen →\rightarrow SPORE)
BA10 (seven unseen PT labels)0.6032 →\rightarrow 0.70830.6289 →\rightarrow 0.7603
Smart-Seq2 (cross-platform)0.4773 →\rightarrow 0.51410.7585 →\rightarrow 0.7821

Together, these results show that SPORE's representation gains are not solely attributable to hierarchy-compatible scoring. Moreover, our original direct-inference audit (Appendix E.3) distinguishes exact-label errors from descendant-compatible predictions and quantifies ontology displacement, revealing differences in prediction specificity and error severity that hAcc alone cannot capture.

The additional evaluations are included in revised Appendices E.1.1 and F.4.

Q4 — Figure 4 readability.

We thank the reviewer for identifying the ambiguity in the Figure 4 legend–panel associations. We have reorganized Figure 4 to clarify the legend–panel association: each row now has a shared legend placed directly below it, and each legend explicitly lists the panels to which it applies ("Biological compartment (A, C, E, F)" and "Ontology relation (B, D)"). Panel numbering and all corresponding references have been updated accordingly.

Reviewer_1zv1OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SPORE, a semantic-aware global post-training framework for improving the reusability of pretrained single-cell foundation models. SPORE incorporates three sources of biological supervision: gene-semantic alignment using GenePT embeddings, ontology-aware cell-type supervision through an asymmetric soft-label objective, and biological-context alignment using BioLORD embeddings. The framework employs a two-stage training procedure to produce a single checkpoint that can be reused across downstream tasks without further backbone adaptation. The authors evaluate SPORE on five foundation model backbones across cell annotation, cancer detection, drug sensitivity, gene-level prediction, regulatory network inference, and perturbation prediction. The experiments demonstrate improvements in several downstream tasks, particularly cell-type annotation, and investigate representation geometry, label efficiency, and direct annotation on unseen studies.

Critical strengths
  • Practical and reusable post-training framework.
  • Broad empirical evaluation.
Critical weaknesses
  • Limited evidence for the benefit of joint semantic supervision. The ablation results in Table 3 suggest that ontology-aware cell supervision accounts for most of the improvements in the strongest tasks. The full SPORE objective achieves nearly identical cell-type accuracy (0.822 vs. 0.821) and cancer AUROC (0.743 vs. 0.742) compared with ontology-only supervision. Moreover, individual supervision branches outperform the joint model on drug sensitivity and tissue specificity. These results weaken the justification for the proposed multi-level integration, which is a central methodological contribution.

  • Generalization evidence and experimental robustness remain limited. The main cell-type transfer benchmark uses the same source corpus for global post-training and downstream readout fitting, although donors are held out for evaluation. This makes it difficult to distinguish generalizable representation improvements from benefits of supervised adaptation to the evaluation domain. While the study-disjoint annotation experiment partially addresses this concern, its label space is restricted to known classes, and the improvement over the strongest competing annotation baseline is relatively small. Furthermore, the main transfer results lack systematic uncertainty estimates across training seeds, making the reliability of smaller improvements difficult to assess.

Questions for authors
  • Could the authors evaluate the ontology-only post-trained model using the direct annotation protocol on held-out studies? This comparison would help establish whether gene-semantic and context supervision improve generalization beyond ontology-aware cell supervision alone.

  • Could the authors provide additional evidence of study-disjoint frozen-feature transfer, or clarify how the existing evaluation distinguishes representation reuse from supervised adaptation to the source corpus? Paired confidence intervals or seed-level variability for the main improvements would also help assess their reliability.

5 comments
AuthorsOpenReview ↗

Official Response to Reviewer 1zv1

Thank you for the constructive feedback on joint supervision and the evidence needed to establish reusable representations. To clarify the contributions of joint supervision, we added leave-one-out controls and an ontology-only comparison on held-out studies. To assess generalization, we added study-disjoint frozen-feature transfer, unseen-label evaluations, and cross-platform evaluations. Reliability is further assessed through independent post-training runs and paired study-bootstrap confidence intervals.

The response is organized into three parts. Part 1 addresses joint supervision (W1/Q1), Part 2 examines representation reuse and generalization (W2/Q2), and Part 3 covers supervised baselines and reliability (W2/Q2).

Part 1 — Joint supervision and OOD generalization (W1, Q1)

1.1 Joint Supervision: Rationale and Component Contributions (W1)

We thank the reviewer for recognizing the practical value of reusable post-training and for highlighting the need to better justify joint semantic supervision.

SPORE asks whether heterogeneous biological supervision can improve a pretrained scFM once through global post-training, yielding a checkpoint reusable across studies and tasks. We develop a framework that decouples semantic supervision from backbone pretraining and integrates gene semantics, ontology-aware cell identity, and biological context through tailored supervision objectives, drawing on corresponding knowledge sources explored in prior work (sciLaMA, scCello, and TEDDY).

We agree that ontology-only supervision accounts for much of the original cell-type and cancer improvements, while individual branches can excel on particular tasks (Table 3). To further characterize these task-dependent contributions, we conducted Geneformer leave-one-out controls, removing each objective while retaining the other two (Appendix E.4.1, Table A19):

ConditionCell hAccCancer AUROCTissue accuracyUGRN AUPRC
Full0.8220.7430.8140.675
−Cell0.7710.7210.8100.677
−Context0.8070.7410.7970.676
−Gene0.8100.7440.7830.675

Removing Cell substantially reduces cell-type and cancer performance, while removing Gene or Context reduces tissue accuracy by approximately 3.1 and 1.7 percentage points, respectively. These results highlight distinct, task-dependent contributions within the joint objective, consistent with the task-specific strengths observed in Table 3. Complementary gradient diagnostics reveal weak positive Cell–Context alignment and nearly orthogonal Gene-related gradients on average, highlighting heterogeneous local interactions among the objectives (Figure A6).

We further evaluated the benefit of joint supervision on independent studies (Section 1.2, below). Full improves direct-annotation macro-F1 over ontology-only post-training from 0.3343 to 0.3658, providing evidence that multi-level semantic integration can enhance cross-study generalization within a reusable global post-training framework.

1.2 Ontology-Only vs. Full on Held-Out Studies (Q1)

We thank the reviewer for suggesting this direct comparison.

To directly address the reviewer's request, we compared ontology-only and Full Geneformer on the same 10,500 cells from 11 study-disjoint datasets, using identical label mappings and direct annotation protocols without target-study fitting:

MetricOntology-onlyFull SPOREΔ\Delta (95% paired CI)
macro-F1 ↑\uparrow0.33430.3658+0.0315 [+0.0008, +0.0492]
accuracy ↑\uparrow0.40060.4292+0.0287 [−0.0215-0.0215, +0.0862]

Using 2,000 paired study-level bootstrap resamples, the improvement in macro-F1 is statistically significant (95% CI strictly above zero), while accuracy exhibits a positive trend despite the CI crossing zero. These results provide evidence that combining semantic and ontology supervision improves annotation across unseen studies.

AuthorsOpenReview ↗

Part 2 — Representation reuse and generalization (W2, Q2)

We next examine representation reuse and generalization across independent studies, unseen labels, and assay technologies (Sections 2.1–2.2), followed by comparisons with supervised annotation baselines and analyses of reliability (Sections 2.3–2.4).

2.1 Study-Disjoint Frozen-Feature Transfer (W2, Q2)

To complement the original donor-held-out transfer (Table 1) and study-disjoint direct annotation (Table 2), we evaluated study-disjoint frozen-feature transfer for cell type annotation using the original Frozen and SPORE Geneformer checkpoints. All 11 evaluation studies are disjoint from global post-training train/validation data by cell, dataset, collection, and DOI. Both backbones remain frozen, with identical MLP readouts trained on separate studies. The closed-set evaluation covers 6,671 cells across 10 evaluable studies.

Under a matched 20-epoch readout protocol, we evaluated three independent readout seeds (13, 37, and 42), using the final-epoch checkpoint for both representations. The following results are equal-study means ±\pm SD across seeds:

MetricFrozenSPOREImprovement
Standard macro-F10.4675 ±\pm 0.00310.5633 ±\pm 0.0064+0.0958
hAcc0.6119 ±\pm 0.00150.7228 ±\pm 0.0042+0.1109
hMacro-F10.5949 ±\pm 0.00170.7147 ±\pm 0.0057+0.1197
Strict accuracy0.4934 ±\pm 0.00280.5704 ±\pm 0.0061+0.0769

These results provide direct evidence that global post-training improves frozen-feature transfer to independent studies, under both conventional and hierarchy-compatible metrics (Appendix E.1.1, Table A6).

Table 1 also provides complementary evidence of cross-task reuse of the same frozen SPORE checkpoint on separate downstream benchmarks, including Cancer-Finder and drug sensitivity, neither of which is a direct post-training objective. The new cross-study evaluation extends this evidence, supporting representation improvements beyond adaptation to the original source corpus.

2.2 Generalization Beyond Known Labels and Assay Technologies (W2)

We thank the reviewer for emphasizing generalization beyond known labels and assay technologies.

To address the known-class restriction, we evaluated the original Frozen and SPORE Geneformer representations on a study-disjoint BA10 developmental-brain cohort comprising six donors and seven cell-type labels absent from direct global post-training supervision.

Frozen-feature classification and retrieval. We assessed donor-held-out classification and cross-donor embedding retrieval:

EvaluationFrozenSPORE
Classification: standard macro-F10.6032 ±\pm 0.00140.7083 ±\pm 0.0034
Classification: accuracy0.6289 ±\pm 0.00400.7603 ±\pm 0.0050
Retrieval: class-macro mAP0.31460.4985
Retrieval: Precision@100.51730.6370
Retrieval: Precision@500.47070.6042

Classification improves across all 18 donor–seed comparisons (six donors ×\times three readout seeds). Retrieval uses fixed cosine similarity without target-label fitting and improves across all seven cell types and six donors (random-ranking mAP: 0.1346), including the three least abundant cell types. These results demonstrate improved classification and direct retrieval of identities absent from post-training supervision (Appendix E.1.1, Tables A7–A8 and Figures A3–A4).

Cross-platform transfer. On an independent Smart-Seq2 fetal hematopoiesis cohort, donor-held-out standard macro-F1 improves from 0.4773 to 0.5141, and accuracy from 0.7585 to 0.7821, supporting representation transfer beyond the 10x platform used for post-training (Appendix F.4, Table A26).

AuthorsOpenReview ↗

Part 3 — Supervised baselines and reliability (W2, Q2)

We now turn to the comparison with supervised annotation baselines and the reliability of the observed improvements.

2.3 Comparison with Strong Supervised Annotation Baselines (W2)

We thank the reviewer for highlighting this comparison. CellTypist, scTab, and HCE-scTab are dedicated supervised cell annotation methods, whereas SPORE globally post-trains existing scFMs to obtain reusable representations across studies and tasks.

Although the OOD hAcc margin over HCE-scTab is modest (0.662 vs. 0.653), SPORE–TF-Sapiens achieves stronger strict accuracy (0.431 vs. 0.318), hierarchical F1 (0.856 vs. 0.786), and lower ontology distance (LCAE: 2.958 vs. 3.944) (Table 2; Appendix E.3). These results reveal advantages in exact-label agreement, ancestor overlap, and ontology displacement that are not captured by hAcc alone.

Importantly, SPORE supports competitive direct annotation without target-study fitting, while the same globally post-trained checkpoint also enables frozen-feature transfer across studies and tasks. This demonstrates the value of direct annotation as part of a broader reusable representation framework.

2.4 Uncertainty and Reproducibility (Q2)

Global post-training reproducibility. Three independent complete Geneformer post-training runs (seeds 37, 38, and 39) yield closely matched scores on cell-type transfer and Cancer-Finder under the same task-specific evaluation protocols (Appendix E.4.3, Table A22). Values are shown as mean ±\pm sample standard deviation (SD) across three independent seeds (n=3).

Global PT seedCell type hAccCell type hMacro-F1Cancer-Finder AUROC
370.82150.46480.7429
380.82170.46350.7402
390.82390.47190.7440
Mean ±\pm SD0.8224 ±\pm 0.00130.4667 ±\pm 0.00450.7423 ±\pm 0.0019

All three runs consistently outperform the corresponding Frozen Geneformer baseline on cell-type hAcc (0.649) and Cancer-Finder AUROC (0.684), with low variability across PT seeds. The consistent Cancer-Finder gains further support the reproducibility of downstream transfer beyond the directly supervised cell-type task.

Paired uncertainty and downstream robustness. On the 11-study direct-annotation cohort, Full SPORE improves standard macro-F1 over matched Frozen+AOSL by +0.0946 [95% paired study-bootstrap CI: 0.0461, 0.1107]. The study-disjoint frozen-feature transfer results in Section 2.1 (Appendix E.1.1) further support the reproducibility of downstream readout evaluation, as all three readout seeds improve over the corresponding frozen baseline across all four metrics.

These analyses provide direct seed-level evidence for the reproducibility of the main Geneformer transfer improvements, complemented by paired study-level uncertainty estimates and separate assessments of downstream readout variability.

Reviewer_MapoOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SPORE, a global post-training framework for improving reusable single-cell foundation model representations. It combines cell ontology supervision, biological-context alignment using BioLORD, and gene-semantic alignment using GenePT. The authors evaluate SPORE across five foundation models and multiple downstream tasks, reporting improvements in cell annotation and selected transfer benchmarks.

Critical strengths
  1. Post training scFMs is an interesting and important direction, especially given that most research / computational resources currently focus on pretraining.
  2. The results demonstrate promising improvements across several downstream tasks and multiple foundation model architectures
Critical weaknesses
  1. The necessity of jointly optimizing three semantic objectives is not supported by the ablations. While Table 3 compares each objective against the full model, it lacks leave-one-out ablations that remove each objective while retaining the other two. This is particularly important because ontology-only supervision nearly matches full SPORE on cell-type and cancer type prediction, and context-only supervision performs better on drug sensitivity and tissue specificity.
  2. My last point also suggests potential interference between objectives. The current paper does not investigate gradient conflicts in multi-task learning.
  3. What's the reason for scGPT and scFoundation (Table 1) not being applied with SPORE? These two models have learnable gene representations, output cell embeddings, and perturbation prediction recipes provided by authors, which arguably fit SPORE and current eval tasks better than the other baselines (eg, scCello or UCE).
  4. For CxG, much of the cell type annotation is very coarse, eg a large portion of cell types is just annotated as "Neuron", which may provide limited supervision for distinguishing fine-grained biological identities. Related, what's the reason for only constructing the post-training dataset from only 10x assays (line 884)? This suggests limited techinical/assay variations in the curated dataset.
  5. One important motivation for scFM pretraining is they can discover biological structure from pretraining data with minimal human biases so that the model could discover novel / rare cell types (such as the Norm cell case study in UCE and the cell line model searching in SCimilarity). Does SPORE post training preserve the ability to distinguish unseen cell types and with in cell type variations? The current evaluations largely focus on known annotation classes and do not address this potential trade off.
Questions for authors

See my weaknesses

Some broader questions that I would be interested in hearing the authors' perspectives on, although they would not affect my evaluation:

  1. Taking an analogy to LLM post training, SFT and RL can improve (and sometimes unlock new) capabilities such as instruction following and long-horizon agentic behavior that are not explicitly optimized during pretraining. In contrast, most tasks evaluated here are already reasonably well supported by pretrained scFMs (Table 1). What new biological capabilities could scFM post training further unlock, beyond improving performance on existing self-supervised models?
  2. Another important aspect of LLM post training is the use of curated, high-quality data rather than simply more pretraining data. How important do the authors think data quality, annotation resolution, and experimental diversity are for scFM post training?
5 comments
AuthorsOpenReview ↗

Official Comment to Reviewer Mapo

Thank you for recognizing the value of post-training and for identifying comparisons that help clarify its effects. We provide additional evidence on joint supervision through objective-removal controls, gradient-interaction diagnostics, and an ontology-only comparison on held-out studies. We then address backbone coverage and evaluate generalization beyond directly supervised labels and the training assay family. Finally, we discuss potential biological applications and the role of supervision quality and diversity.

Part 1 — Joint supervision and backbone coverage

W1–W2 — Joint supervision and optimization interactions

We thank the reviewer for suggesting leave-one-out controls and gradient-interaction analysis.

Joint supervision (W1). SPORE asks whether a pretrained scFM can be improved through one-time global post-training with heterogeneous biological supervision, yielding representations reusable across studies and tasks. Motivated by prior work leveraging gene-level knowledge (sciLaMA), cell ontologies (scCello), and broader biological annotations (TEDDY), we integrate these three complementary sources of semantic supervision.

We further evaluated Geneformer leave-one-out controls, removing one objective at a time while retaining the other two, using validation-selected checkpoints (Appendix E.4.1, Table A19 and Figure A6A).

ConditionCell hAccCancer AUROCTissue accuracyUGRN AUPRC
Full0.8220.7430.8140.675
−Cell0.7710.7210.8100.677
−Context0.8070.7410.7970.676
−Gene0.8100.7440.7830.675

Cell supervision contributes most strongly to cell-type and cancer prediction, while removing Gene or Context reduces tissue accuracy by approximately 3.1 and 1.7 percentage points, respectively. These task-dependent contributions motivate their integration into a single reusable checkpoint.

Gradient interactions (W2). We examined objective gradients across 32 fixed training batches at the Full checkpoint, without parameter updates (Appendix E.4.1, Figure A6B). Cell–Context gradients show weak positive mean alignment (cosine +0.103; 8/32 negative batches), while gradients involving Gene are nearly orthogonal on average, with mixed directions (18/32 negative batches for each pair). These diagnostics reveal heterogeneous local interactions, without evidence of strong, systematic gradient opposition at the evaluated checkpoint.

Cross-study generalization. On 11 unseen studies (10,500 cells), Full Geneformer improves direct-annotation macro-F1 over ontology-only post-training from 0.3343 to 0.3658 (paired study-bootstrap Δ\Delta +0.0315, 95% CI [0.0008, 0.0492]; Appendix E.3, Table A18). This provides evidence that combining semantic and ontology supervision improves annotation generalization across unseen studies.

W3 — Backbone coverage.

We appreciate the reviewer’s suggestion to consider broader backbone coverage. We selected the current backbones to prioritize diversity in pretraining objectives and biological priors. Specifically, Geneformer represents generic masked-gene self-supervision; scCello incorporates cell-ontology structure during pretraining; and TEDDY-G leverages broader biological annotations. UCE and TF-Sapiens use different-scale protein-derived gene embeddings (ESM-2 15B and 3B, respectively) and differ in their pretraining objectives (masked-expression prediction vs. autoregressive modeling). Given computational constraints, we prioritized this diversity over an exhaustive backbone sweep.

We agree that scGPT and scFoundation are valuable additional candidates, offering binned-expression and continuous-value encoding architectures, respectively. Both were included as frozen baselines in Table 1, rather than excluded due to architectural incompatibility. Extending paired post-training evaluations to these models requires additional model-specific integration and validation, and would further broaden the scope of our study.

The backbone-selection scope is described in Appendix C.2.

AuthorsOpenReview ↗

Part 2 — Annotation resolution and generalization

W4 — Annotation resolution and assay choice.

We appreciate the reviewer’s emphasis on annotation resolution and assay diversity.

Annotation resolution. We selected CELLxGENE Census as one of the most comprehensive curated resources for human single-cell transcriptomics, offering large-scale coverage and harmonized Cell Ontology annotations. Our filtered corpus includes annotations at different resolutions. For example, 3,004 training cells carry the generic "neuron" label, while 8,954 cells have 53 more specific neuronal labels (74.88% of the neuronal lineage). To accommodate such variation, SPORE supervises cell identity through a prediction head without explicitly constraining embedding geometry, while AOSL uses ontology-aware soft labels to account for different annotation levels. As more finely annotated atlases become available, incorporating them should enable richer semantic supervision.

Assay selection. Following scTab (Fischer et al., 2024), we selected the predominant 10x assay family in CELLxGENE for global post-training. Its scale provides sufficient training data, while more consistent measurement protocols help limit technical confounding during supervised learning.

We additionally evaluated the original Frozen and SPORE Geneformer checkpoints on Ranzoni et al.'s Smart-Seq2 fetal hematopoiesis cohort (2,070 cells). Both frozen-feature classification and trajectory conservation improved:

EvaluationFrozenSPORE
Classification: standard macro-F10.47730.5141
Classification: accuracy0.75850.7821
Trajectory: 11-sample macro0.56330.5690
Trajectory: all cells pooled0.69110.7118

These results support cell-identity transfer and preservation of continuous developmental structure on a non-10x platform. Complete protocols and results are provided in the revised appendix.

The added Smart-Seq2 protocols and results are in Appendix F.4, Tables A26–A27 and Figure A10.

W5 — Types outside the PT vocabulary and biological variation.

We agree that scFM representations should remain useful for exploring biological identities beyond those explicitly supervised during post-training. To examine this, we evaluated the original Frozen and SPORE Geneformer representations on a study-disjoint BA10 developmental-brain cohort comprising 15,171 cells from six donors. All seven evaluated cell-type labels are absent from global post-training supervision.

Frozen-feature classification. With both backbones frozen, donor-held-out MLP readouts improve standard macro-F1 from 0.6032 ±\pm 0.0014 to 0.7083 ±\pm 0.0034 across three readout seeds. All 18 paired donor–seed comparisons improve (six donors ×\times three seeds).

Cross-donor retrieval. We further evaluated direct cell retrieval using the same frozen embeddings, with fixed cosine similarity and candidates restricted to other donors. No target-label fitting was performed. The seven-class macro results are:

MetricFrozenSPORE
mAP0.31460.4985
Precision@100.51730.6370
Precision@500.47070.6042

The random-ranking mAP is 0.1346. Retrieval improves across all seven cell types and all six donors. The three least abundant cell types in this cohort also show improvements:

Cell typeCellsFrozen mAPSPORE mAP
Forebrain radial glial cell3570.0850.120
Committed oligodendrocyte precursor3960.3630.829
MGE-derived cortical interneuron3380.1350.529

These results show that SPORE preserves and improves the ability to distinguish and retrieve cell identities beyond its direct post-training labels, including less abundant populations. Continuous-trajectory evaluations provide complementary evidence: the sample-macro conservation score improves from 0.7212 to 0.7668 on Oetjen erythroid differentiation and from 0.5633 to 0.5690 on Ranzoni Smart-Seq2 fetal hematopoiesis (Appendices F.3–F.4, Tables A25 and A27).

The added BA10 classification and retrieval results are in Appendix E.1.1, Tables A7–A8 and Figures A3–A4.

AuthorsOpenReview ↗

Part 3 — Future capabilities and supervision quality

Response to Q1 (Potential new biological capabilities).

We see global post-training as a step toward plug-and-play scFMs, making biological representations more reusable and task-relevant information more accessible across studies and tasks. SPORE provides an initial example through a reusable checkpoint supporting frozen transfer across tasks (Table 1), direct annotation on study-disjoint data (Section 4.3), and label-efficient downstream prediction (Section 4.2).

Looking ahead, we envision biological instruction following and compositional generalization as particularly promising next steps. By incorporating textual knowledge, ontologies, and structured instructions, future post-training could enable models to retrieve, compare, and analyze cells according to user-specified biological questions, including unseen combinations of cell identities, tissues, and conditions, without developing separate predictors for each task. Further extensions toward RL-style post-training could support multi-step biological analysis and hypothesis generation, provided that suitable verifiable rewards and benchmarks are developed.

Response to Q2 (Data quality, annotation resolution, and experimental diversity).

We appreciate the reviewer's emphasis on data curation. Reliable and consistent annotations are particularly important for scFM post-training. SPORE partially addresses these considerations through data filtering and ontology-aware supervision, while more reliable and fine-grained annotations could provide richer supervision.

Beyond data quality, we consider biological and experimental diversity equally important. In particular, covering diverse cell identity ×\times tissue ×\times condition combinations may be more valuable for generalization than adding more cells from already well-represented combinations. Diversity across sequencing technologies, platforms, and experimental protocols may further help models learn representations robust to technical variation. We therefore see the quality and diversity of semantic supervision, rather than raw cell count alone, as a key consideration for future scFM post-training.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Paper Revision

We sincerely thank all reviewers for their constructive feedback. In response, we have revised the manuscript, with substantive changes highlighted in blue. The major updates are summarized below.

  1. Joint Supervision. We added leave-one-objective-out ablations and gradient-interaction diagnostics (Table A19; Figure A6), together with an ontology-only comparison on 11 held-out studies (Table A18). These analyses clarify the task-dependent contributions of the three semantic objectives and their effects on cross-study generalization, as further discussed in Sections 4.3–4.4. (Reviewers Mapo, 1zv1, and jPTP)

  2. Representation Reuse and Generalization. We expanded the evaluation with (i) study-disjoint frozen-feature transfer across three readout seeds (Table A6), (ii) classification and cross-donor retrieval for seven cell types absent from global post-training supervision (Tables A7–A8; Figures A3–A4), and (iii) cross-platform classification on Smart-Seq2 data (Table A26), providing complementary evidence for representation reuse across independent studies, cell types absent from direct global post-training supervision, and assay technologies (see also Section 4.1 for the cross-study findings). (Reviewers Mapo, 1zv1, and jPTP)

  3. Preservation of Continuous Structure. We evaluated trajectory conservation on two developmental datasets (Tables A25 and A27; Figures A9–A10), examining whether semantic post-training preserves continuous biological variation beyond supervised cell identities; both evaluations show maintained or improved trajectory conservation (Section 4.2). (Reviewers Mapo and jPTP)

  4. Evaluation Metrics and Reproducibility. We expanded the cell-type transfer evaluation with conventional exact-match accuracy, macro-F1, and standardized hMacro-F1 reporting in the 375-class benchmark (Table A5). We additionally report three-seed post-training reproducibility (Table A22) and paired study-bootstrap confidence intervals for direct annotation (Table A18). (Reviewers 1zv1 and jPTP)

  5. Presentation and Readability. We improved the legend–panel associations and caption of Figure 4 and updated the Conclusion to reflect the additional generalization evidence. (Reviewers Mapo and jPTP)

We believe these additional experiments and clarifications have strengthened the manuscript, and we look forward to your further feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.