acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Feature Propagation with Reciprocal Neighbors for Single-Cell Imputation

Abstract

Graph-based feature propagation (FP) provides an efficient, non-parametric approach to single-cell RNA sequencing (scRNA-seq) imputation. However, severe sparsity distorts cell neighbor rankings and causes over-smoothing when using static graphs and fixed propagation depths. We propose scRFP, a training-free FP framework that adapts neighbor selection, propagation depth, and graph degree without requiring cell-type labels. Rather than using reciprocal cell pairs directly as sparse graph edges, scRFP employs them to derive a closed-form metric transformation that downweights directions that vary within reciprocal pairs and recalibrates neighbor similarities across all cells. We then monitor neighborhood contrast to stop hard propagation, and prune low-confidence edges to accommodate varying cellular densities before soft diffusion. Across seven benchmark scRNA-seq datasets with a single configuration, scRFP achieves the best clustering performance on six, improving ARI by up to 12.1% (relative) over the strongest baseline, and the lowest technical-dropout recovery error in most settings, at runtime comparable to existing FP methods; removing any of the three adaptations lowers average ARI.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
scRNA-seq, Clustering, Feature Propagation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_3uBROpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

scRFP is an untrained propagation method for sparse scRNA-seq data. It uses reciprocal neighbors to adjust cell similarities, adaptively stops hard propagation, and prunes the neighborhood before soft propagation. The method is evaluated on seven public datasets.

Critical strengths

The method has a simple design and does not require neural network training. Its three components target different parts of propagation: neighbor selection, propagation depth, and neighborhood size. It achieves the best clustering results on six of the seven datasets, and the ablations support the usefulness of these components.

Critical weaknesses

The hard-propagation stage uses adaptive stopping, but the soft-propagation stage always runs for 80 steps. The paper does not clearly explain this difference or show how sensitive the results are to this setting. The method is described as label-free, but the stopping threshold is selected using ARI on the other six datasets. This means that labels are still used during parameter selection, and the claim should be stated more precisely.

Questions for authors

Why is soft propagation fixed at 80 steps after the adaptive hard-propagation stage? Please provide a brief justification or sensitivity result.

Which labels are used during parameter selection, and which part of the method is truly label-free?

2 comments

We sincerely thank the reviewer for the positive assessment and the thoughtful questions, which helped us state the scope of our claims more precisely.


[W1] Biological zeros and expression leakage

We agree with the reviewer that masking nonzero entries tests the recovery of technical dropout rather than the preservation of biological zeros. As an additional analysis, we call a cell type negative for a gene if fewer than 1% of its cells have a nonzero raw count, and positive if at least 50% do. Negative types must contain at least 30 cells. Genes with both a negative and a positive type give 1,122 gene–dataset combinations. These populations are defined from detection rates and are not independently validated biological negatives. AUROC measures how well a method ranks positive cells above negative cells. Leak is the percentage of negative cells whose value exceeds the lower quartile of positive-cell values for the same gene and method. Both metrics are computed per gene, averaged within each dataset, and then averaged equally across datasets. Cell-type labels are used only for this evaluation. We compare scRFP with the three FP methods, which share its propagation framework.

scRFPscFPscBFPscCR
AUROC ↑.989.948.973.996
Leak (%) ↓0.675.273.000.33

scRFP has a Leak of 0.67%, compared with 5.27% for scFP, 3.00% for scBFP, and 0.33% for scCR. This analysis measures relative separation in empirically defined low-detection populations rather than preservation of genuine biological zeros.


[W2/Q1] Label use in parameter selection and baseline settings

We thank the reviewer for this comment and agree that our label-free description requires qualification. scRFP uses no cell-type labels during imputation, but labels are used for ARI-based parameter selection.

Appendix F.2 reports leave-one-dataset-out selection of the stopping threshold τ. For each held-out dataset, τ is chosen from the five tested values by maximizing mean ARI on the other six datasets. The held-out labels are used only to evaluate the selected τ. All seven splits select τ = 0.00745, the shared default used in Table 1. This exclusion of target labels applies only to τ. L = 80 was selected using clustering performance on all seven datasets, including each target dataset ([W2/Q2]). We will distinguish label use in parameter selection from the absence of label use during imputation.

For the baselines, we used the authors' released implementations with their default hyperparameters and no additional tuning, under the same preprocessing and evaluation protocol. We will state explicitly that this protocol differs from our development-stage parameter selection. Our comparison covers nine baselines including the recent propagation-based methods scFP, scBFP, and scCR.


[W2] Reciprocal edges without the global transformation

We thank the reviewer for this suggestion. We ran this comparison in response to Reviewer BE6y [W2/Q2], where the full per-dataset table is given. Using reciprocal edges directly as the initial graph (mutual kNN), with all other steps of scRFP unchanged, gives an average ARI of .602. Adding each cell’s nearest neighbor raises this to .660, compared with .830 for scRFP. These results support using reciprocal pairs to estimate a shared transformation over directly using them as graph edges within the same propagation pipeline.


[W2/Q2] Choice of L = 80 and sensitivity to L and the initial neighbor count

The adaptive stopping rule controls hard propagation before graph reconstruction. Soft propagation then operates on the resulting fixed graph for a shared number of steps L. We set k = 15 and α = 0.99 following scFP and scCR, and selected L = 80 empirically based on clustering performance across the seven benchmark datasets. The table below reports clustering ARI when varying the number of soft-propagation steps L and the initial neighbor count k. We vary one parameter at a time and keep all other settings at their defaults.

SettingBaron HumanBaron MouseMouse BladderPancreasWorm NeuronZeiselPBMC 10kAverage
Default (L = 80, k = 15).900.907.790.957.667.874.717.830
L = 20.798.711.697.957.676.880.626.763
L = 40 (scFP/scBFP/scCR default).813.648.698.959.675.877.611.754
k = 5.814.876.542.934.605.880.603.750
k = 10 (scBFP default).824.918.652.947.571.871.646.775

L = 80 and k = 15 achieve the highest average ARI among the tested settings, although individual datasets favor different values. We will add these sensitivity results and clarify the empirical choice of L = 80 in the paper.


We hope these responses address the reviewer's questions, and we would be happy to provide any further clarification.

AuthorsOpenReview ↗

Kind Follow-Up from Authors

Dear Reviewer 3uBR,

As the rebuttal period draws to a close, we wanted to follow up on our response to your questions. Thank you again for your thoughtful comments.

Regarding the choice of 80 steps, we clarified that adaptive stopping determines when to end hard propagation before graph reconstruction. Soft propagation then runs on the reconstructed graph for a fixed number of steps shared across datasets. We selected 80 steps empirically. Across the seven datasets, 80 steps gave a mean clustering ARI of .830, compared with .763 for 20 steps and .754 for 40 steps.

We also clarified how we used cell-type labels. For each held-out dataset, we selected the stopping threshold τ\tau using clustering ARI on the other six datasets. All seven splits selected the same τ\tau value. We selected the soft-propagation depth using clustering ARI on all seven datasets, including the target dataset. Once these parameters are set, scRFP performs imputation without cell-type labels.

We would appreciate it if you could take a look at our response. We hope these clarifications address your concerns, and we would be happy to answer any further questions.

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

This paper proposes a method called scRFP for single-cell RNA sequencing (scRNA-seq) imputation. It uses reciprocal nearest neighbors to improve cell similarity estimation, adaptively determines when to stop feature propagation, and adjusts the number of neighbors for each cell. The method does not require model training or cell-type labels. Empirical evaluations are conducted on seven benchmark datasets using clustering and dropout recovery tasks.

Critical strengths
  1. Strong experimental results across multiple datasets.

The proposed method achieves the best clustering performance on six out of seven datasets. It also shows good results in dropout recovery. These consistent improvements over existing methods are a strong point of the paper.

  1. An efficient method without model training.

The method does not require neural network training or cell-type labels. It also uses the same hyperparameter settings across all seven datasets. This makes the method simple and practical.

Critical weaknesses
  1. The effectiveness of adaptive stopping is not clearly demonstrated.

The paper claims that adaptive stopping helps prevent over-smoothing. However, the stopping criterion reaches the maximum of 300 iterations on three datasets. Also, the ablation study only compares adaptive stopping with a fixed depth of 20 iterations. Therefore, it is unclear whether the improvement comes from the proposed stopping rule or simply from using more propagation steps. This weakens one of the main contributions of the paper.

  1. Limited novelty and insufficient comparisons with simpler alternatives.

The proposed method combines existing ideas, such as reciprocal nearest neighbors and metric learning, with adaptive propagation. Although the results are strong, it is unclear whether the proposed components are better than simpler alternatives. For example, the paper does not compare its similarity transformation with standard metric learning methods. Also, the adaptive neighborhood method is not compared with a fixed-degree graph that has a similar average number of neighbors. These missing comparisons make it difficult to judge the novelty and effectiveness of the proposed components.

Questions for authors
  1. Clarification: In Figure 8, the stopping criterion reaches the maximum of 300 iterations on three datasets. Does this mean that adaptive stopping does not actually stop propagation early on these datasets? How does this support the claim that the method prevents over-smoothing?

  2. Additional evidence: Could the authors compare adaptive stopping with several fixed propagation depths, such as 20, 80, 150, and 300? This comparison would help show whether the stopping rule itself improves performance.

  3. Additional evidence: Could the authors compare the reciprocal-pair similarity transformation with simpler alternatives, such as PCA whitening or standard metric learning methods? Also, could they compare adaptive neighborhoods with a fixed-degree graph using a similar average number of neighbors? These comparisons would help clarify the contributions of the proposed components.

3 comments

We sincerely thank the reviewer for the careful review and for proposing specific controls.


[W1/Q1] Early stopping and the iteration cap

Yes, when hard propagation reaches tmax=300t_{max} = 300, the criterion has not triggered early stopping, and propagation ends at the cap. These cases therefore do not provide evidence of an early-stopping benefit.

On the three datasets where the criterion stops before the cap, the selected depths yield higher clustering ARI than continuing to 300 steps, with all other settings unchanged.

DatasetSelected depthAdaptive stopping (ours)Fixed 300
Mouse Bladder19.790.684
Worm Neuron95.667.539
PBMC 10k2.717.572

Increasing the cap to 500 or 1000 leaves these early-stopped results unchanged and changes the overall mean ARI from .830 to .834 and .833, respectively. These results show that adaptive stopping can avoid the loss in clustering quality caused by prolonged propagation on these datasets. We will clarify that it does not guarantee prevention of over-smoothing on every dataset.


[W1/Q2] Comparison with fixed propagation depths

We varied only the hard-propagation depth and kept all other settings unchanged. The table reports mean clustering ARI over the seven datasets.

Hard-propagation depthMean ARI
Adaptive (ours).830
Fixed 20.787
Fixed 40.792
Fixed 80.785
Fixed 150.781
Fixed 300.776

Adaptive stopping achieves a mean ARI of .830, exceeding the best shared fixed depth tested, 40 steps (.792). Increasing the fixed depth to 80, 150, or 300 does not reproduce this performance, indicating that simply running more propagation steps does not explain the adaptive result.


[W2/Q3] Comparison with simpler transformations and matched-degree graphs

We replaced one component at a time, keeping the remaining pipeline and hyperparameters unchanged. The table reports mean clustering ARI across seven datasets.

ComponentAlternativeMean ARI
Full methodscRFP.830
Initial graphPCA-50 with cosine kNN.691
Initial graphPCA-50 whitening with cosine kNN.651
Initial graphRCA within-pair whitening, 2 rounds.765
Soft-propagation neighborhoodFixed k matched to our mean degree per dataset.703

For initial graph construction, our transformation achieves higher mean ARI than untransformed PCA, PCA whitening, and RCA within-pair whitening. Unlike RCA whitening, our rescaling accounts for total variance relative to within-pair variation. For soft propagation, matching each dataset’s mean neighborhood size with a fixed k yields .703, compared with .830 for adaptive neighborhoods. Thus, matching the average number of neighbors does not reproduce the adaptive result. We will add the full per-dataset comparisons and implementation details for these controls to the appendix.


[W2] Novelty

Reciprocal neighbors and metric learning have established precedents. A common use of reciprocal neighbors is a mutual-kNN graph, which leaves cells with few reciprocal neighbors poorly connected. scRFP instead uses reciprocal pairs to estimate a transformation shared across all cells, so every cell receives a neighbor list in the transformed space. Unlike RCA, which whitens the within-pair scatter, scRFP rescales each direction by the square root of the ratio of its total variance to its within-pair scatter. This emphasizes directions with high overall variance relative to their variation within reciprocal pairs. We use a shared stopping threshold and percentile setting across datasets, while propagation depth and neighborhood sizes adapt to each input. We will clarify these differences in the revised paper.


We thank the reviewer again and hope these additional analyses address the concerns raised. We would be happy to provide any further clarification.

AuthorsOpenReview ↗

Kind Follow-Up from Authors

Dear Reviewer iiZK,

As the rebuttal period is approaching its end, we wanted to kindly follow up on our responses to your valuable comments and questions.

To address your concerns, we clarified that runs reaching tmax⁡=300t_{\max} = 300 do not provide evidence of early stopping. On the datasets where the criterion stops early, we compared the selected depths with continued propagation to 300 steps, and we also tested larger values of tmax⁡t_{\max} (500500 and 1,0001,000). We further compared adaptive stopping with fixed hard-propagation depths, including those you suggested. We also compared our reciprocal-pair transformation with untransformed PCA, PCA whitening, and RCA, and compared adaptive neighborhoods with fixed-size neighborhoods matched to the mean degree of the adaptive neighborhoods in each dataset. Finally, we clarified how scRFP differs from mutual-kNN graphs and from RCA.

We would greatly appreciate it if you could review our responses. We hope that these additional analyses address the concerns you raised. Should you have any further questions, we would be more than happy to address them.

I thank the authors for their detailed responses. However, the explanation of methodological novelty is not sufficiently convincing. Thus, my overall assessment remains unchanged.

Reviewer_BE6yOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

It proposes scRFP, a training-free method for single-cell RNA-seq imputation built on the two-stage feature propagation framework of scFP. It introduces a reciprocal-neighbor-based metric transformation for initial graph construction, a neighborhood-contrast criterion for stopping hard propagation, and similarity-based neighbor selection before soft propagation.

Critical strengths
  1. The empirical improvements over existing FP methods are fairly consistent. scRFP achieves the best ARI on six of seven datasets and also performs well on masked-expression recovery.

  2. The method is straightforward and does not require neural network training. The reported average runtime is comparable to scFP, with lower peak GPU memory usage, making the improvements potentially useful in practice.

Critical weaknesses
  1. The motivation would be stronger with a direct analysis of how the neighborhood problem evolves through the existing scFP pipeline across datasets. Figure 1 shows cross-type neighbors selected by cosine kNN, but it is unclear which representation is used. Since scFP already performs hard propagation and then rebuilds the graph before soft propagation, poor neighborhoods in the input expression alone do not establish how much of this problem remains after its refinement. A stage-by-stage comparison of neighborhood quality for scFP and scRFP would help show what the proposed components actually address.

  2. I find the methodological advance over scFP somewhat limited. The hard/soft propagation framework remains unchanged, while the additions mainly concern metric adjustment, stopping, and neighbor selection. The reciprocal-pair transformation is a meaningful addition, but its advantage over simpler graph refinement alternatives is not sufficiently established. The evaluation also focuses on relatively small datasets, leaving the practical scope of the improvements unclear. In particular, repeated neighbor searches during hard propagation warrant validation on a larger dataset, such as the Macosko mouse retina dataset.

Questions for authors
  1. Which representation is used to construct the neighborhoods in Figure 1: the input expression, the hard-propagated expression, or the final imputed expression? Can you compare cross-type neighbor fractions at the corresponding stages of scFP and scRFP to show whether scFP's graph reconstruction already alleviates the initial problem?

  2. How does the reciprocal-pair metric transformation compare with simpler alternatives, such as directly using a mutual-kNN graph? Such a comparison would help assess whether the metric transformation itself is necessary for the observed gains.

  3. Have you evaluated scRFP on a larger dataset, such as the Macosko mouse retina dataset? Both recovery quality and runtime/memory would be useful, particularly because the stopping criterion requires repeated neighbor searches.

4 comments

We sincerely thank the reviewer for the thorough and constructive review. We address each point below.


[W1/Q1] Figure 1 clarification and stage-wise neighborhood comparison

Figure 1 uses the input expression: the log-normalized 2,000-gene matrix that scFP and scRFP take as input, with cosine top-15 neighbors (the cell itself excluded). We will state this in the caption.

To compare the two pipelines stage by stage, we applied the measure of Figure 1(b) at each stage: the percentage of a cell's top-15 neighbors whose annotation differs from its own, averaged over the seven datasets. scFP's stages were reproduced with the authors' code.

Top-15 neighbor listscFPscRFP
Cosine, input expression (Figure 1(b))17.2%17.2%
Initial neighbor list used to build the hard-propagation graph17.2% (cosine, input expression)7.0% (cosine, reciprocal-pair rescaled PCA space)
Cosine, hard-propagated expression (scFP rebuilds its graph from this list)10.2%6.8%
Cosine, final imputed expression9.2%6.9%

For scRFP, the third row evaluates cosine top-15 neighbors before the adaptive neighborhood selection used for soft propagation. scFP's graph reconstruction clearly reduces cross-type neighbors (17.2% → 10.2%). However, scRFP retains a lower cross-type neighbor fraction both after hard propagation (scRFP 6.8% vs. scFP 10.2%) and in the final imputed expression (scRFP 6.9% vs. scFP 9.2%). scRFP's initial neighbor list already has a cross-type neighbor fraction of 7.0% before propagation, showing that its improved neighborhood separation is present from the outset. We will add this table to the paper.


[W2/Q2] Comparison with mutual-kNN graphs

Following the reviewer's suggestion, we replaced our initial graph with mutual-kNN graphs and kept all other steps of scRFP unchanged. Two cells are connected only if each is among the other's 15 cosine nearest neighbors on the input expression. Because many cells then have no other neighbor, we also tested adding an edge from each cell to its nearest neighbor (Mutual kNN + 1-NN). Cosine kNN on the input, the initial graph of scFP, is included as a reference. All values are clustering ARI.

Initial graphBaron HumanBaron MouseMouse BladderPancreasWorm NeuronZeiselPBMC 10kAverage
Cosine kNN after the reciprocal-pair transformation (ours).900.907.790.957.667.874.717.830
Cosine kNN on the input (scFP's initial graph).820.665.559.955.451.845.545.691
Mutual kNN.792.608.459.830.399.458.671.602
Mutual kNN + 1-NN.677.709.652.915.414.695.561.660

Our initial graph achieves the highest ARI on all seven datasets. The mutual-kNN graph yields an average ARI of .602, compared with .830 for our method and .691 for cosine kNN on the input. On the input, 42–93% of cells have no non-self mutual neighbor and receive no information from other cells during hard propagation. Adding a nearest-neighbor edge for every cell eliminates isolated nodes and raises the average to .660, but this does not close the gap. These results support using reciprocal pairs to estimate a shared metric rather than restricting graph edges to reciprocal pairs.


[W2] Methodological advance over scFP

We retain scFP's hard/soft propagation updates. Our methodological contribution lies in adapting which neighbors supply information, how long hard propagation runs, and how many neighbors are retained for soft propagation. Controlled ablations within the same framework show that each component contributes to the average ARI. Replacing our initial graph with cosine kNN on the input expression reduces the average ARI from .830 to .691. Using a single fixed hard-propagation depth across datasets achieves at most .792 among the depths tested. A fixed neighbor count matched to our mean degree on each dataset yields .703.

[W2/Q3] Evaluation on a larger dataset: PBMC 68k (68,579 cells)

We thank the reviewer for this suggestion. We evaluated scRFP on PBMC 68k (68,579 cells). The MNAR recovery and clustering protocols are identical to Section 5. Runtime and memory were measured on one NVIDIA A100 80 GB.

scRFPscFPscBFPscCR
Clustering ARI.231.230.250.229
MNAR recovery RMSE (rate 0.2).241.332.375.686
Wall-clock time per run, s48OOM1071OOM
Peak GPU memory, GB4.9OOM1.3OOM

Clustering scores are averaged over 10 seeds for all methods. Recovery errors are averaged over 10 seeds for scRFP and scBFP and over 3 seeds for scFP and scCR. Runtime and peak memory refer to GPU execution. scFP and scCR ran out of memory on an NVIDIA A100 80 GB GPU because their released implementations construct the kNN graph using dense 68,579 × 68,579 matrices. For these two methods, recovery and clustering results were therefore obtained using CPU execution.

scRFP achieves the lowest recovery error under MNAR masking, at .241 against .332 for scFP, .375 for scBFP, and .686 for scCR. In clustering, scRFP has an ARI similar to scFP and scCR, and scBFP obtains the highest scores on all three metrics. scRFP completes each run in 48 s, approximately 22× faster than scBFP, with a peak GPU memory usage of 4.9 GB.

On this dataset, the stopping criterion selects T = 3, resulting in seven neighbor searches in total. Each search computes top-k neighbors in chunks on the GPU without materializing the full 68,579 × 68,579 similarity matrix. The reported runtime includes these searches.


We hope these additional results address the reviewer's concerns, and we would be grateful if the reviewer would take them into account in the final assessment. We would be happy to answer any further questions.

Thank you for the rebuttal. The stage-wise neighborhood comparison and mutual-kNN controls address my concerns, and the PBMC 68k results provide useful evidence of scalability. I still think the methodological advance remains incremental and clustering gains on the larger dataset are limited, the added evidence strengthens the empirical contribution. I am happy to raise my rating to weak acceptance.

Thank you for your updated evaluation! We sincerely appreciate your thoughtful feedback and constructive suggestions, which have helped us improve our paper. We will incorporate the additional experiments, clarifications, and discussions into the revised version. We are grateful for your recognition of our contributions and truly appreciate your positive assessment of our work. Should you have any further questions or concerns, we would be happy to address them.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their constructive feedback. We are encouraged that the reviewers found scRFP simple and training-free (BE6y, iiZK, 3uBR), consistently strong in the clustering task(BE6y, iiZK, 3uBR), and effective in dropout recovery (BE6y, iiZK). The reviewers also noted its practical runtime and memory usage (BE6y), its single hyperparameter setting across all datasets (iiZK), and ablations that support each of its three components (3uBR). We have revised the paper, with all changes marked in blue. The revisions are summarized below.

Main text

  • Figure 1 (caption): specifies that the neighborhoods are built on the input expression. (BE6y)
  • Section 4.1: clarifies how our rescaling differs from Relevant Component Analysis. (iiZK)
  • Sections 4.2 and 5.3: revise the description of adaptive stopping. (iiZK)
  • Section 5.4, Table 5: comparison with mutual-kNN graphs as alternative initial graphs. (BE6y, 3uBR)
  • Section 5.4, Table 6: results on a larger dataset, PBMC 68k (68,579 cells). (BE6y)

Appendix

  • Appendix B: baseline settings. (3uBR)
  • Appendix C: distinguishes label use in parameter selection from label-free imputation. (3uBR)
  • Appendix F.2, Tables 14–15: sensitivity to the soft-propagation steps LL and the initial neighbor kk. (3uBR)
  • Appendix F.3, Table 16: comparison with PCA whitening, RCA whitening, and a degree-matched fixed kk. (iiZK)
  • Appendix F.3: average ARI with fixed hard-propagation depths of 20 to 300 steps. (iiZK)
  • Appendix F.4, Table 17: stage-wise cross-type neighbor fraction of scFP and scRFP. (BE6y)

Detailed answers are given in our individual responses. We hope that these revisions address the reviewers' questions and concerns. We thank the reviewers again for their time and thoughtful reviews, which have helped us improve the paper, and we would be happy to answer any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.