acceptodds
Under review as a conference paper at ICLR 2027

RankFuse: Integrating single-cell representations through neighbour rankings

Abstract

No single-cell foundation model is best on every dataset, yet their embeddings live in incompatible coordinate systems, so their complementary strengths cannot be pooled. We present RankFuse, an encoder-agnostic method that combines frozen single-cell foundation models through what they agree on: each cell's ranked nearest neighbours across batches. Reciprocal rank fusion merges the per-encoder rankings into a single graph, and a one-layer graph convolution learns an integrated representation for batch integration, with a variance term to prevent collapse and a batch adversary to remove batch effects. No encoder is fine-tuned, so the adapter trains a small fraction of the parameters of any contributing model and new encoders can be added at the cost of a neighbour search. On PBMC and an immune cell atlas, fusing scGPT, Geneformer, a batch-adversarial scVI and a DINO encoder, both pretrained on a 46-study corpus, improves the scIB benchmark score over the strongest single encoder by up to 5.6% relative, and yields the strongest batch correction on the atlas. Ablations show that fusion beats a single-encoder graph with the same adapter, and that performance depends far more on which encoders build the graph than on which supply node features, suggesting that neighbour-rank agreement is a coordinate-free interface for combining representation models, of which single-cell encoders are one instance.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single-cell integration, rank fusion, 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_tsGhOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes RankFuse, a method for integrating heterogeneous single-cell encoders through their cross-batch nearest-neighbor rankings. It combines these rankings using reciprocal rank fusion, constructs a unified graph, and trains a lightweight graph adapter over whitened encoder features. Experiments on three single-cell datasets investigate integration performance and the respective contributions of graph construction, feature selection, and adaptation

Critical strengths

The method provides a simple and practical interface for combining heterogeneous encoders without requiring explicit coordinate alignment. The experimental analysis usefully separates graph construction from feature adaptation and includes single-encoder controls and component ablations. The paper also transparently reports dataset-dependent gains and trade-offs between batch correction and biological conservation, which helps clarify the scope of its conclusions.

Critical weaknesses

My main concern is that the motivation presents differences in embedding coordinates and scales as a stronger obstacle to direct fusion than the evidence supports. Incompatible coordinates prevent naive coordinate-wise averaging, but do not inherently preclude concatenation or a learned fusion module. Although the paper evaluates whitened concatenation, this relatively limited control does not establish that rank-based fusion is preferable to direct feature fusion with comparable adaptation capacity. A matched concatenation baseline with a learned projection or lightweight fusion network would help determine whether the gains arise specifically from the proposed rank-based interface rather than from the graph construction and subsequent adaptation

Questions for authors
  1. Could you clarify which forms of direct fusion are considered problematic and distinguish coordinate-wise averaging from concatenation or learned feature fusion? 2. Can you compare RankFuse with whitened concatenation followed by a trainable projection or lightweight fusion network under comparable training conditions and adaptation capacity? 3. Can you better isolate the contribution of rank fusion from graph-based adaptation, for example by comparing alternative graph constructions while holding the features and adapter fixed? I would consider raising my score if the authors can convincingly address this concern and demonstrate the benefit of the proposed fusion interface through matched comparisons.
Reviewer_Rqr7OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RankFuse, a framework for integrating heterogeneous single-cell representations through cross-batch neighbor rankings. Reciprocal rank fusion produces a fixed graph, and a single graph-convolutional adapter processes whitened, concatenated encoder embeddings using VISReg and a batch-adversarial objective. Experiments on three datasets consider fixed-checkpoint and target-trained encoder pools. The fixed-checkpoint fusion improves scIB Total over the strongest contributing encoder on PBMC and the immune atlas, but not pancreas. Controls show that the contributions of fusion and learned adaptation vary across datasets and encoder settings.

Critical strengths
  1. A coherent framework for reusing heterogeneous encoders. Neighbor rankings provide a shared interface without requiring alignment of embedding coordinates. The separation between graph and feature rosters offers a useful way to combine relational information and representation features while keeping the contributing encoders frozen.

  2. Meaningful controls support some benefits beyond learned adaptation. On the fixed-checkpoint atlas experiment, untrained fusion already exceeds the strongest contributing encoder, and the full method exceeds adapted scGPT by 0.035 across all three adapter seeds. Together with the fixed-propagation and roster controls, these results help distinguish the roles of fusion and adaptation. The authors also report negative results and limitations explicitly.

Critical weaknesses

The evidence for preferring fusion over the strongest individually adapted encoder remains limited. The atlas result against adapted scGPT is encouraging. However, Section 4.2 explicitly limits the atlas and pancreas controls to adapted scGPT, rather than establishing the strongest adapted checkpoint on those datasets. On PBMC, where scGPT is the strongest adapted single-checkpoint arm, the 0.012 fusion margin falls within the authors’ prespecified inconclusive band; on pancreas, fusion and adapted scGPT perform similarly. These results support benefits in particular comparisons, but leave uncertainty about when combining multiple encoders is preferable to adapting a strong individual encoder.

Questions for authors

Section 4.2 reports adapting all four fixed checkpoints individually on PBMC. Were equivalent comparisons performed on the atlas and pancreas? If so, how does RankFuse compare with the strongest individually adapted checkpoint on each dataset?

Reviewer_L6FYOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RankFuse integrates representations from heterogeneous single-cell encoders by combining their cross-batch neighbour rankings through reciprocal rank fusion. The resulting graph is fixed, while whitened encoder embeddings provide features for a single-layer graph convolutional adapter trained with VISReg and a batch-adversarial objective. The paper evaluates fixed-checkpoint and target-trained encoder pools on three datasets. Fixed-checkpoint fusion improves scIB Total over the strongest contributing checkpoint on PBMC and the immune atlas, whereas the target-trained results vary across datasets. The supplementary material provides implementation details, matched single-encoder controls, roster-selection comparisons, and analyses of the trained layer’s benefits and biological-conservation costs.

Critical strengths

1.The method provides a common interface for combining heterogeneous encoders through neighbour rankings, without fine-tuning the contributing encoders. Its integration adapter has 1,640 parameters, with a separate training-only discriminator of approximately 6,300 parameters. In the fixed-checkpoint setting, RankFuse increases scIB Total from the strongest contributing checkpoint’s 0.625 to 0.660 on PBMC and from 0.597 to 0.609 on the immune atlas. These results support the practical potential of combining existing encoder checkpoints. 2.The paper uses matched single-encoder adaptation, degree-preserving graph shuffling, and crossed graph/feature rosters to investigate the sources of improvement. The supplement also directly compares minimum-overlap selection against choosing the four strongest encoders on PBMC: Total is 0.7061 versus 0.6720 across three seeds, with standard deviations of 0.0100 and 0.0020, respectively. This provides empirical support for the selection rule in that comparison. The authors also report that minimum overlap loses on neighbourhood identity, appropriately exposing the trade-off rather than reporting only the favourable aggregate.

Critical weaknesses

1.The target-trained comparison uses cell-type labels from the subsequently evaluated cells to determine early stopping for the four author-trained batch-aware models. The reference encoder used in roster selection is also chosen through a label-derived identity score. The eight self-supervised baselines instead use fixed training durations without this label exposure. Although the adapter itself uses a fixed 3,000-step schedule and does not use cell-type labels in its objective, the overall target-trained pipeline still depends on these labels through encoder training and reference selection. The reported comparison therefore does not establish performance under a common label-independent selection protocol. 2.The matched controls do not establish a consistent advantage over the best equivalently adapted single encoder across datasets. On PBMC, the fusion margin is 0.012, within the authors’ unresolved band; on pancreas, the margin over adapted scGPT is negligible. The atlas advantage of 0.035 is encouraging, but the supplement explicitly limits that comparison to adapted scGPT rather than the best adapted contributing encoder. The trained layer also has a measurable cost: Table S4 shows that it decreases Bio and increases Batch in all nine target-trained seed pairs, while decreasing Total on all three atlas seeds by 0.004–0.012. These findings leave the benefit of the complete pipeline insufficiently established across the reported settings.

Questions for authors

Q1.Performance without evaluation-label-assisted selection When the four author-trained encoders and the reference-encoder selection are evaluated under a common protocol that does not use the evaluated cell-type labels for stopping or selection, do the main target-trained conclusions remain unchanged, particularly the PBMC advantage reported in Table 3? Q2.The best matched single-encoder baseline on atlas and pancreas On the atlas and pancreas, if each of the four contributing checkpoints receives its own cross-batch graph constructed using the same procedure and the same adapter training protocol, what is RankFuse’s paired Total difference from the best adapted single encoder, measured in the same scoring pass? Q3.Reselecting rosters after excluding Concerto Section S.6.1 states that the pipeline includes a low-signal encoder and that excluding it has not been tested. If Concerto is removed from the target-trained pool and the rosters are reselected using the same rule, with the remaining pipeline unchanged, how do Total, Bio, and Identity change relative to the reported configuration?

Comments on the paper

2 comments
Area_Chair_sZEWOpenReview ↗

Reminder - Please review authors' rebuttal

Dear Reviewers,

A quick update on the discussion: we are now halfway through the discussion period, and the authors have posted their rebuttal.

I would appreciate it if you could take some time to join the discussion at this stage. In particular, please go through the authors' responses, consider whether the concerns and questions you raised have been addressed, and let the authors know if you have any further feedback for them.

Getting everyone's input now will help us move toward a well-grounded final decision. Many thanks for your participation.

Best regards,

Your AC

AuthorsOpenReview ↗

Letter to AC

Dear AC sZEW,

Thank you for putting up the reminder and overseeing this review process, we understand the reviewers and ACs are under considerable amount of pressure these days thus would like to express our gratitude here.

The problem we set out to address is chooing among pretrained single-cell RNA encoders, spanning from classic scVI and its variants to transformer based scGPT models, especially when cell-type labels are unavailable, which is phenomenon in most of public data. RankFuse is developed combining the four frozen checkpoints we tested without retraining the encoders.

In response to L6FY and Rqr7, we compared RankFuse with each checkpoint given its own cross-batch graph and the same adaptation procedure (Appendix S.12). Against adapted scGPT, the strongest adapted checkpoint, the atlas gain is still considerable and passed our own selection gate. As for smaller, less available dataset PBMC, the result remains undetermined, and we cannot establish an advantage on pancreas, where the data geometry could be different or limited by the random seed variation, which is out of the scope of this study. We have revised our abstract to reflect these findings.

Reviewer L6FY also identified two uses of cell-type labels in the separate setting where encoders are trained on the target data. After removing both, RankFuse ranks second on PBMC, so we have narrowed that claim and documented where cell-type labels enter the workflow. We appreciate the comments as it indeed helped us to test the robustness of our method, however we would kindly note here, this additional analysis meant to examine the separate case where the encoders themselves are trained on the target data.

Lastly, reviewer tsGh questioned about direct feature fusion. The cross-batch graph does improve coverage compared with the projection. We'd like to thank reviewer tsGh raising this question as the result shows that rank fusion does not uniquely responsible for the coverage gain, which is largely anticipated, and it does support the practical value of our graph-based approach.

We see RankFuse as a practical way for labs to combine existing checkpoints when they cannot use cell-type labels to choose among them. The revision makes the evidence for that use, as well as its limitations, more explicit.

Thank you and all the reviewers for considering our work.

Best regards, RankFuse Authors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.