acceptodds
Under review as a conference paper at ICLR 2027

ABSORB: Absorbing-State Discrete Diffusion for Single-Cell Gene Expression Imputation

Abstract

Single-cell RNA sequencing produces matrices that are highly sparse, yet a large fraction of zero entries are missing: either they are silent in the gene expression or are lost during transcript capture. This ambiguity makes imputation a selective missing-data problem in contrast to a generic denoising problem. We introduce ABSORB, an absorbing-state discrete diffusion process that represent gene expressions using a dedicated absorbing [MASK] state. Our proposed framework tokenizes the expression into per-gene equal-mass quantile bins, and learns to recover masked observed entries from the remaining expression profile. During inference, measured entries remain fixed while candidate dropouts are replaced with a [MASK] and reconstructed by a role-conditioned Transformer. The model predicts categorical distributions for each gene rather than a conditional mean, ensuring non-negativity is held by construction. To distinguish genuine recovery of missing expression from the sparse input, we evaluate only against known truth and external labels. Across a four-level simulated dataset spanning 50–72% dropout, ABSORB achieves Spearman correlations of – and log-scale RMSE of –, outperforming nine baselines. Additionally, we evaluate cell-type clustering as a label-assisted downstream analysis, while separating it from the independent evidence provided by held-out gene expression recovery.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single-cell RNA sequencing, gene expression imputation, discrete diffusion

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_uRbhOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ABSORB, which formulates the imputation task as a "selective missing-data" problem and performs selective imputation of scRNA-seq data using an absorbing-state discrete diffusion framework. The paper quantifies numerical recovery at held-out positions with known ground truth, and compares against nine baselines on simulated datasets at four different dropout rates as well as multiple real datasets. The results show that ABSORB has a clear advantage on simulated data, and this advantage widens further as the dropout rate increases.

Critical strengths

This paper frames scRNA-seq imputation as a "selective missing-data" problem and, based on absorbing-state discrete diffusion, realizes selective imputation that "keeps observed entries fixed and reconstructs only candidate dropouts." The problem formulation is clear and clearly distinguished from the continuous-diffusion denoising route. Under a controlled held-out evaluation, the method performs well on multiple recovery metrics.

Critical weaknesses

(1) The paper does not experimentally decompose the contribution of each component, nor does it try alternative decoding architectures (e.g., a conditional autoencoder or an MLP decoder). It is therefore difficult to judge whether the performance comes primarily from absorbing-state diffusion itself or from masked token prediction, the Transformer, tokenization, or confidence-first decoding; although the supplementary material provides a theoretical analysis of the role/timestep embeddings, this is not sufficient to replace an experimental ablation.

(2) ABSORB imputes only candidate dropouts with a mask value of 0, and does not handle "non-zero but unreliable" observations whose counts are deflated by dropout; at the evaluation level, it scores only at held-out positions with known ground truth, whereas most non-diffusion baselines impute the full matrix (scIDPMs and scImpute are themselves selective imputers). Under this setting, imputation performance depends on the reliability of the candidate-dropout mask, yet the current held-out evaluation is not sufficient to comprehensively measure the consequences of mask misclassification—especially the cost of erroneously imputing biological zeros. Its overall imputation utility in real scenarios therefore still requires further validation.

Questions for authors

(1) The current method determines whether a gene is expressed in a cluster solely based on whether that gene appears non-zero at least once in the cluster: once a non-zero appears within the cluster, its zeros are treated as candidate dropouts. This criterion is simple and easily confounded by sequencing contamination, ambient RNA, doublets, batch effects, and the like. It may both misclassify a contaminated biological zero as a dropout (false positive) and retain a genuine dropout that is zero throughout the cluster as a biological zero (false negative). The reasonableness of this assumption deserves further discussion.

(2) The paper mainly constructs dropout through simulated data or by down sampling real data; it remains unclear whether, when applied to real datasets without artificially simulated dropout, the method can recover genuine biological information (such as cell-type marker genes, cell-type proportions, and gene correlations)

10 comments
AuthorsOpenReview ↗

Clarifications on component contribution

We thank the reviewer for the constructive comments and for recognizing our contribution along with the distinction between selective missing-data reconstruction and continuous denoising.

Component contribution…

We thank the reviewer for raising this important point. To address this we have added a matched masked-Transformer ablation using the same backbone and tokenization. ABSORB performs better in 64/70 dataset–metric comparisons, including 14/14 RMSE comparisons, providing empirical evidence that the gains are not explained by tokenization and masked Transformer prediction alone. We have added this ablation study to the Appendix section 2.9.

Our novelty lies in the combination of the following design choices:

  • Selective absorbing state: Unlike approaches that introduce Gaussian noise or transitions between numerical count states, ABSORB represents missing expressions using a dedicated [MASK] token. This separates uncertainty about missing expression from the underlying expression values, allowing the model to learn the recovery of masked entries through discrete-state prediction.
  • Immutable measured context: Throughout reconstruction, experimentally observed non-zero expression values remain fixed, while only candidate dropout positions are reconstructed. As the model iteratively imputes missing entries, the measured expression values continue to provide a reliable conditioning context rather than being modified by successive predictions. This differs from full-matrix denoising or smoothing approaches, such as DCA and MAGIC, which also modify observed expression values.
  • Gene-specific ordinal representation: Genes naturally exhibit different expression distributions, with some being highly expressed and others having consistently lower expression levels. Rather than applying a shared discretization scale across all genes, ABSORB constructs gene-specific quantile bins based on each gene's positive expression distribution, preserving zero as a separate state. This allows expression levels to be represented relative to each gene's own distribution, providing more balanced resolution for both highly and weakly expressed genes.
  • Confidence-first iterative reconstruction: High-confidence predictions are committed first and subsequently used as context for reconstructing the remaining masked entries.

Together, these choices distinguish ABSORB from Gaussian diffusion approaches such as scIDPMs and count-space diffusion with binomial thinning in methods such as CountsDiff, against which ABSORB also demonstrates stronger recovery performance.

AuthorsOpenReview ↗

Evidence of ablation study on ABSORB vs. Masked-MLP Transformer

Matched ablation: ABSORB vs. masked-Transformer controls trained at fixed mask rates of 0.50 and 0.30 (columns .50 and .30), with the same backbone and tokenization. Held-out recovery; bold marks the best of the three.

Datasetρ↑\rho \uparrow ABSORB / .50 / .30RMSE ↓\downarrow ABSORB / .50 / .30ED ↓\downarrow ABSORB / .50 / .30MMD ↓\downarrow ABSORB / .50 / .30SW ↓\downarrow ABSORB / .50 / .30
Sim D10.922 / 0.907 / 0.9080.609 / 0.675 / 0.667362 / 449 / 4380.144 / 0.125 / 0.11129.7 / 32.7 / 30.5
Sim D20.928 / 0.904 / 0.9100.595 / 0.697 / 0.678268 / 305 / 2970.110 / 0.0997 / 0.092726.4 / 26.5 / 26.7
Sim D30.918 / 0.906 / 0.9100.634 / 0.670 / 0.652220 / 274 / 2810.0826 / 0.0840 / 0.084523.5 / 25.2 / 25.0
Sim D40.942 / 0.917 / 0.9220.549 / 0.670 / 0.644153 / 196 / 2410.0552 / 0.0568 / 0.065320.0 / 22.3 / 24.7
PBMC68k MCAR500.197 / 0.209 / 0.2140.478 / 0.485 / 0.4960.125 / 0.117 / 0.1130.00260 / 0.00199 / 0.001900.0727 / 0.0723 / 0.0671
PBMC68k MNAR250.334 / 0.244 / 0.2340.412 / 0.424 / 0.4840.0517 / 0.0941 / 0.1760.000705 / 0.00169 / 0.003700.0487 / 0.0640 / 0.0931
Bone Marrow MCAR500.392 / 0.324 / 0.3100.703 / 0.834 / 0.7873.37 / 9.45 / 3.900.0209 / 0.0646 / 0.02720.570 / 0.879 / 0.609
Bone Marrow MNAR250.402 / 0.335 / 0.3070.634 / 0.806 / 0.7620.541 / 2.02 / 1.250.00322 / 0.0122 / 0.007560.224 / 0.450 / 0.365
Human Brain MCAR500.280 / 0.239 / 0.2322.543 / 2.638 / 2.731471 / 506 / 5360.0167 / 0.0171 / 0.0175107 / 109 / 110
Human Brain MNAR250.249 / 0.227 / 0.2232.547 / 2.559 / 2.72295.5 / 121 / 2500.00243 / 0.00311 / 0.0072043.5 / 53.3 / 78.5
VISp MCAR500.510 / 0.350 / 0.2972.376 / 2.981 / 2.9281,396 / 2,749 / 3,6780.0147 / 0.0269 / 0.0322153 / 211 / 236
VISp MNAR250.514 / 0.253 / 0.4042.271 / 2.979 / 2.564106 / 535 / 1560.000800 / 0.00451 / 0.0012944.6 / 91.3 / 48.4
Dentate Gyrus MCAR500.481 / 0.464 / 0.4770.709 / 0.949 / 1.09012.8 / 69.3 / 1410.0397 / 0.162 / 0.2600.736 / 2.03 / 3.52
Dentate Gyrus MNAR250.438 / 0.426 / 0.4330.768 / 1.140 / 0.7983.73 / 36.8 / 7.410.0116 / 0.104 / 0.02360.421 / 1.67 / 0.612
ABSORB better13/1414/1413/1411/1413/14
AuthorsOpenReview ↗

Clarifications on candidate-mask reliability and biological zeros

Candidate-mask reliability and biological zeros…

We thank the reviewer for highlighting this concern. We agree that identifying potential dropouts based on the presence of at least one non-zero value within a cluster can be sensitive to noise and may lead to incorrect imputation of biological zeros. In the original submission, we reported results at α = 0, as this operating point prioritizes maximum dropout recovery. However, the selectivity parameter α allows us to control this trade-off by defining a dataset-adaptive expression-frequency threshold based on the α-quantile of expression frequencies associated with observed non-zero entries. Candidate imputations for genes expressed below this threshold within their cluster are retained as zeros, thereby reducing the influence of rare, potentially noisy expression observations. As shown in our sensitivity analysis, α = 0.05 substantially reduces erroneous biological-zero imputation while maintaining high genuine-dropout recovery. Nevertheless, we acknowledge that distinguishing biological zeros from technical dropouts remains challenging, particularly under contamination or severe sparsity, and have included this analysis and the associated limitations in the Appendix section 2.6.

ABSORB at α = 0 is the original submission; the zero-decision rows use α = 0.02 and α = 0.05.

MethodBiological zeros imputed ↓\downarrowReal dropouts recovered ↑\uparrow
ABSORB, α = 0 (original submission)65.5%91.7%
ABSORB + zero decision, α = 0.0235.9%91.1%
ABSORB + zero decision, α = 0.0512.6%90.2%
SAVER0.1%26.5%
ALRA2.4%72.6%
MAGIC3.4%84.6%
scImpute3.5%81.4%
DCA21.6%83.7%
countsdiff32.8%65.4%
DeepImpute91.3%94.3%

Held-out recovery before → after the zero decision (α = 0 → α = 0.05) on the real datasets. Bold marks the better value; unbolded pairs are ties at the reported precision.

Datasetρ↑\rho \uparrowRMSE ↓\downarrowED ↓\downarrowMMD ↓\downarrowSW ↓\downarrow
PBMC68k MCAR500.197 → 0.1230.478 → 0.4870.125 → 0.1780.00260 → 0.004150.0727 → 0.0868
PBMC68k MNAR250.334 → 0.2820.412 → 0.4230.0517 → 0.05140.000705 → 0.0006710.0487 → 0.0485
Human Brain MCAR500.280 → 0.2982.543 → 2.568471 → 4670.0167 → 0.0166107 → 107
Human Brain MNAR250.249 → 0.2782.547 → 2.56295.5 → 93.10.00243 → 0.0023543.5 → 43.1
Bone Marrow MCAR500.392 → 0.4150.703 → 0.7073.37 → 3.290.0209 → 0.02050.570 → 0.565
Bone Marrow MNAR250.402 → 0.4280.634 → 0.6380.541 → 0.5060.00322 → 0.003010.224 → 0.214
VISp MCAR500.510 → 0.5342.376 → 2.3851,396 → 1,3950.0147 → 0.0146153 → 153
VISp MNAR250.514 → 0.5412.271 → 2.305106 → 1030.000800 → 0.00077244.6 → 44.2
Dentate Gyrus MCAR500.481 → 0.5000.709 → 0.71412.8 → 12.60.0397 → 0.03900.736 → 0.727
Dentate Gyrus MNAR250.438 → 0.4600.768 → 0.7693.73 → 3.590.0116 → 0.01110.421 → 0.410
Zero decision better8/100/109/109/107/10 (2 ties)
Reviewer_8wobOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces ABSORB, an absorbing-state discrete diffusion framework for selective imputation of scRNA-seq data. It combines a cell-group-based candidate-dropout mask, per-gene quantile tokenization, and a Transformer trained to reconstruct masked observed entries. At inference, confidence-first decoding progressively fills candidate missing entries while preserving observed values. The framework is evaluated through simulated-data recovery, artificially masked real-data recovery, and label-assisted clustering analyses.

Critical strengths
  1. Coherent formulation for selective imputation. The framework explicitly represents unknown entries with an absorbing mask state and preserves observed values during imputation. Its combination of quantile tokenization and conditional reconstruction is well motivated for sparse, long-tailed data and provides a practical adaptation of discrete diffusion to this task.

  2. Consistent reconstruction performance. The evaluation includes known-truth simulations and held-out recovery on real datasets under both random and value-dependent masking. ABSORB achieves the lowest log-RMSE in all six real-data settings reported in Figure 5. This consistency provides meaningful evidence of its numerical recovery capability, although its advantages are less uniform on other metrics.

Critical weaknesses
  1. The risks of imputation are not fully discussed. Imputation is known to introduce spurious structure, such as false gene–gene correlations and false differential expression (e.g., Andrews & Hemberg, 2018). Zero counts in UMI data are also largely consistent with the sampling of lowly expressed genes (Svensson, 2020; Sarkar & Stephens, 2021). Beyond using the latter point to motivate selective imputation, the paper does not discuss whether ABSORB avoids these problems, nor does it evaluate false signals. Its evaluation cannot detect them, because recovery is scored only at entries with positive true values, so imputing a true zero is never penalised.

  2. Limited novelty and missing ablations. The main modules are pre-existing: absorbing-state discrete diffusion (D3PM), confidence-based parallel decoding (MaskGIT), and masked prediction of binned expression values, as used in single-cell models such as scBERT and scGPT. Masked discrete diffusion has also been applied to single-cell data (ScDiVa). The authors state that the contribution lies in combining these modules. However, only the number of quantile bins is ablated. Without ablations of the decoding procedure and the mask construction, it is unclear which design choices account for the reported gains.

Questions for authors
  1. Can the authors evaluate whether ABSORB introduces spurious signals? On the simulated data, where the true matrix is available, how do gene–gene correlations and between-group differential expression after imputation compare with those of the true matrix, relative to the baselines?

  2. Since recovery is scored only at positions with positive true values, what fraction of true-zero positions does ABSORB impute as non-zero on the simulated data, and how does this compare with the baselines?

  3. Can the authors provide ablations of the main design choices? Examples include confidence-first iterative decoding versus a matched one-pass predictor and random-order decoding using the same trained backbone.

  4. How were the baselines configured? Please explain the negative Spearman correlations of SAVER and scImpute in Figure 4.

  5. Could the authors compare ABSORB with pretrained single-cell foundation models (e.g., scGPT or scFoundation)?

14 comments
AuthorsOpenReview ↗

Clarifications on novelty and ablation studies

We thank the reviewer for the detailed comments and for recognizing ABSORB's consistent numerical reconstruction performance.

Limited novelty and missing ablation…

We agree that our novelty is not any individual component in isolation, but the following task-specific formulation:

  • Selective absorbing state: Unlike approaches that introduce Gaussian noise or transitions between numerical count states, ABSORB represents missing expression using a dedicated [MASK] token. This separates uncertainty about missing expression from the underlying expression values, allowing the model to learn the recovery of masked entries through discrete-state prediction.
  • Immutable measured context: Throughout reconstruction, experimentally observed non-zero expression values remain fixed, while only candidate dropout positions are reconstructed. As the model iteratively imputes missing entries, the measured expression values continue to provide a reliable conditioning context rather than being modified by successive predictions. This differs from full-matrix denoising or smoothing approaches, such as DCA and MAGIC, which also modify observed expression values.
  • Gene-specific ordinal representation: Genes naturally exhibit different expression distributions, with some being highly expressed and others having consistently lower expression levels. Rather than applying a shared discretization scale across all genes, ABSORB constructs gene-specific quantile bins based on each gene's positive expression distribution, preserving zero as a separate state. This allows expression levels to be represented relative to each gene's own distribution, providing more balanced resolution for both highly and weakly expressed genes.
  • Confidence-first iterative reconstruction: High-confidence predictions are committed first and subsequently used as context for reconstructing the remaining masked entries.

The closest diffusion alternatives use fundamentally different state-space semantics: Gaussian corruption in scIDPMs and count-space diffusion with binomial thinning in CountsDiff.

We agree, however, that the original submission did not experimentally isolate the individual contribution of the iterative formulation. We have therefore added a matched ablation, to isolate the iterative formulation, by comparing ABSORB with matched masked-Transformer controls using the same backbone. ABSORB performs better in 64/70 dataset–metric comparisons, including all 14 RMSE comparisons. We have added this ablation study to the Appendix section 2.9, which allows us to sharpen our contribution rather than attributing all gains generically to “diffusion.”

AuthorsOpenReview ↗

Evidence of ablation study on ABSORB vs. Masked-MLP Transformer

Matched ablation: ABSORB vs. masked-Transformer controls trained at fixed mask rates of 0.50 and 0.30 (columns .50 and .30), with the same backbone and tokenization. Held-out recovery; bold marks the best of the three.

Datasetρ↑\rho \uparrow ABSORB / .50 / .30RMSE ↓\downarrow ABSORB / .50 / .30ED ↓\downarrow ABSORB / .50 / .30MMD ↓\downarrow ABSORB / .50 / .30SW ↓\downarrow ABSORB / .50 / .30
Sim D10.922 / 0.907 / 0.9080.609 / 0.675 / 0.667362 / 449 / 4380.144 / 0.125 / 0.11129.7 / 32.7 / 30.5
Sim D20.928 / 0.904 / 0.9100.595 / 0.697 / 0.678268 / 305 / 2970.110 / 0.0997 / 0.092726.4 / 26.5 / 26.7
Sim D30.918 / 0.906 / 0.9100.634 / 0.670 / 0.652220 / 274 / 2810.0826 / 0.0840 / 0.084523.5 / 25.2 / 25.0
Sim D40.942 / 0.917 / 0.9220.549 / 0.670 / 0.644153 / 196 / 2410.0552 / 0.0568 / 0.065320.0 / 22.3 / 24.7
PBMC68k MCAR500.197 / 0.209 / 0.2140.478 / 0.485 / 0.4960.125 / 0.117 / 0.1130.00260 / 0.00199 / 0.001900.0727 / 0.0723 / 0.0671
PBMC68k MNAR250.334 / 0.244 / 0.2340.412 / 0.424 / 0.4840.0517 / 0.0941 / 0.1760.000705 / 0.00169 / 0.003700.0487 / 0.0640 / 0.0931
Bone Marrow MCAR500.392 / 0.324 / 0.3100.703 / 0.834 / 0.7873.37 / 9.45 / 3.900.0209 / 0.0646 / 0.02720.570 / 0.879 / 0.609
Bone Marrow MNAR250.402 / 0.335 / 0.3070.634 / 0.806 / 0.7620.541 / 2.02 / 1.250.00322 / 0.0122 / 0.007560.224 / 0.450 / 0.365
Human Brain MCAR500.280 / 0.239 / 0.2322.543 / 2.638 / 2.731471 / 506 / 5360.0167 / 0.0171 / 0.0175107 / 109 / 110
Human Brain MNAR250.249 / 0.227 / 0.2232.547 / 2.559 / 2.72295.5 / 121 / 2500.00243 / 0.00311 / 0.0072043.5 / 53.3 / 78.5
VISp MCAR500.510 / 0.350 / 0.2972.376 / 2.981 / 2.9281,396 / 2,749 / 3,6780.0147 / 0.0269 / 0.0322153 / 211 / 236
VISp MNAR250.514 / 0.253 / 0.4042.271 / 2.979 / 2.564106 / 535 / 1560.000800 / 0.00451 / 0.0012944.6 / 91.3 / 48.4
Dentate Gyrus MCAR500.481 / 0.464 / 0.4770.709 / 0.949 / 1.09012.8 / 69.3 / 1410.0397 / 0.162 / 0.2600.736 / 2.03 / 3.52
Dentate Gyrus MNAR250.438 / 0.426 / 0.4330.768 / 1.140 / 0.7983.73 / 36.8 / 7.410.0116 / 0.104 / 0.02360.421 / 1.67 / 0.612
ABSORB better13/1414/1413/1411/1413/14
AuthorsOpenReview ↗

Clarifications on true biological zeros

True biological zeros…

We agree with the reviewer that selective imputation requires balancing the recovery of genuine dropouts against the preservation of biological zeros. We evaluated different values of the selectivity parameter (α), which determines the expression-frequency threshold used to distinguish potential dropouts from biological zeros within cell clusters. Specifically, the threshold is determined from the α-quantile of observed expression frequencies, allowing more conservative imputation as α increases. The original submission reported the α = 0 operating point to prioritize maximum dropout recovery. However, the sensitivity analysis demonstrates a clear, controllable trade-off, where α = 0.05 substantially reduces erroneous biological-zero imputation while maintaining high genuine-dropout recovery, as shown in the table below.

ABSORB at α = 0 is the original submission; the zero-decision rows use α = 0.02 and α = 0.05.

MethodBiological zeros imputed ↓\downarrowReal dropouts recovered ↑\uparrow
ABSORB, α = 0 (original submission)65.5%91.7%
ABSORB + zero decision, α = 0.0235.9%91.1%
ABSORB + zero decision, α = 0.0512.6%90.2%
SAVER0.1%26.5%
ALRA2.4%72.6%
MAGIC3.4%84.6%
scImpute3.5%81.4%
DCA21.6%83.7%
countsdiff32.8%65.4%
DeepImpute91.3%94.3%

Held-out recovery before → after the zero decision (α = 0 → α = 0.05) on the real datasets. Bold marks the better value; unbolded pairs are ties at the reported precision.

Datasetρ↑\rho \uparrowRMSE ↓\downarrowED ↓\downarrowMMD ↓\downarrowSW ↓\downarrow
PBMC68k MCAR500.197 → 0.1230.478 → 0.4870.125 → 0.1780.00260 → 0.004150.0727 → 0.0868
PBMC68k MNAR250.334 → 0.2820.412 → 0.4230.0517 → 0.05140.000705 → 0.0006710.0487 → 0.0485
Human Brain MCAR500.280 → 0.2982.543 → 2.568471 → 4670.0167 → 0.0166107 → 107
Human Brain MNAR250.249 → 0.2782.547 → 2.56295.5 → 93.10.00243 → 0.0023543.5 → 43.1
Bone Marrow MCAR500.392 → 0.4150.703 → 0.7073.37 → 3.290.0209 → 0.02050.570 → 0.565
Bone Marrow MNAR250.402 → 0.4280.634 → 0.6380.541 → 0.5060.00322 → 0.003010.224 → 0.214
VISp MCAR500.510 → 0.5342.376 → 2.3851,396 → 1,3950.0147 → 0.0146153 → 153
VISp MNAR250.514 → 0.5412.271 → 2.305106 → 1030.000800 → 0.00077244.6 → 44.2
Dentate Gyrus MCAR500.481 → 0.5000.709 → 0.71412.8 → 12.60.0397 → 0.03900.736 → 0.727
Dentate Gyrus MNAR250.438 → 0.4600.768 → 0.7693.73 → 3.590.0116 → 0.01110.421 → 0.410
Zero decision better8/100/109/109/107/10 (2 ties)

Versatility of our proposed algorithm under the selectivity parameter (α) is demonstrated in Figure 16 of Appendix section 2.6.

Importantly, this trade-off is not unique to ABSORB. Methods exhibiting stronger biological-zero preservation, such as ALRA, scImpute, and MAGIC, primarily rely on low-rank approximation, statistical modeling, and graph-based smoothing, respectively, but generally achieve lower dropout recovery. In contrast, deep-learning and generative approaches, including DCA, CountsDiff, and DeepImpute, tend to modify a larger proportion of biological zeros in our evaluation. Nevertheless, at α = 0.05, ABSORB achieves a more favorable balance, retaining higher dropout recovery than the zero-preserving baselines while substantially improving biological-zero preservation. We have included this sensitivity analysis and a discussion of the broader selectivity trade-off in the Appendix section 2.6.

Reviewer_CbtSOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a model called ABSORB for single-cell gene expression imputation using absorbing-state discrete diffusion. It identifies potential missing values based on cell-type clusters, converts gene expression values into quantile-based tokens, and trains a Transformer to predict masked values. During inference, it keeps observed values unchanged and gradually fills missing values using confidence-based decoding. Experiments are conducted on simulated and real datasets, comparing the proposed method with nine baseline methods.

Critical strengths
  1. Strong performance in gene expression recovery

The proposed method shows strong results on simulated datasets with different dropout levels. It also performs well on several real datasets. The comparisons with nine baseline methods support the effectiveness of the proposed approach.

  1. A clear approach to handling missing gene expression values

The paper uses a special [MASK] token to represent missing values and keeps observed values unchanged. This design fits the imputation problem well and ensures non-negative predictions. It provides a clear motivation for the proposed framework.

Critical weaknesses
  1. Limited novelty and insufficient ablation studies

The main components, including absorbing-state diffusion, masked prediction, and confidence-based decoding, are based on existing methods. More importantly, the paper does not provide enough ablation studies to show which components are responsible for the performance gains. In particular, it is unclear whether the diffusion process offers a meaningful advantage over a standard masked Transformer.

  1. Insufficient evaluation of selective imputation

The paper aims to recover missing values while preserving true biological zeros. However, the evaluation mainly focuses on recovering masked non-zero values. It does not clearly show whether the method incorrectly changes true biological zeros. In addition, the use of cell-type labels for dropout mask estimation raises concerns about the performance in settings where such labels are unavailable. These issues make it difficult to fully assess the effectiveness of selective imputation.

Questions for authors
  1. What is the contribution of the diffusion process itself? How does ABSORB compare with a standard masked Transformer using the same backbone and tokenization, but without iterative diffusion-based decoding? This comparison would help clarify the main technical contribution.
  2. How well does ABSORB preserve true biological zeros? Can the authors report how often true biological zeros are incorrectly identified as dropouts and replaced with non-zero values? This is important for evaluating the main goal of selective imputation.
  3. How much does the method depend on cell-type labels? Can the authors clarify how the performance changes when the dropout mask is estimated without ground-truth cell-type labels? This would help assess the practical usefulness of the method in unlabeled datasets.
9 comments
AuthorsOpenReview ↗

Clarifications on novelty and ablation studies

We thank the reviewer for the careful assessment and for recognizing ABSORB's strong reconstruction performance and the suitability of an explicit absorbing state for selective imputation.

Limited novelty and insufficient ablation studies…

We acknowledge that absorbing-state diffusion, masked prediction, and confidence-based decoding have been explored in prior work, both in broader machine learning applications and in single-cell modeling. Our contribution lies in their task-specific adaptation and integration into a unified framework for selective scRNA-seq imputation, specifically designed to recover candidate dropouts while preserving observed gene-expression measurements.

More specifically, our novelty lies in the combination of the following design choices:

  • Selective absorbing state: Unlike approaches that introduce Gaussian noise or transitions between numerical count states, ABSORB represents missing expressions using a dedicated [MASK] token. This separates uncertainty about missing expression from the underlying expression values, allowing the model to learn the recovery of masked entries through discrete-state prediction.
  • Immutable measured context: Throughout reconstruction, experimentally observed non-zero expression values remain fixed, while only candidate dropout positions are reconstructed. As the model iteratively imputes missing entries, the measured expression values continue to provide a reliable conditioning context rather than being modified by successive predictions. This differs from full-matrix denoising or smoothing approaches, such as DCA and MAGIC, which also modify observed expression values.
  • Gene-specific ordinal representation: Genes naturally exhibit different expression distributions, with some being highly expressed and others having consistently lower expression levels. Rather than applying a shared discretization scale across all genes, ABSORB constructs gene-specific quantile bins based on each gene's positive expression distribution, preserving zero as a separate state. This allows expression levels to be represented relative to each gene's own distribution, providing more balanced resolution for both highly and weakly expressed genes.
  • Confidence-first iterative reconstruction: High-confidence predictions are committed first and subsequently used as context for reconstructing the remaining masked entries.

We would also like to highlight this difference conceptually from continuous scIDPMs and count-space diffusion with binomial thinning in CountsDiff. ABSORB treats dropout as an explicit [MASK] state representing absence of measurement, rather than as a numerical count evolving toward or from zero. This separates missingness from valid expression states and provides a more natural formulation for selective recovery. The distinction is also empirical, with ABSORB substantially outperforming both diffusion baselines on held-out recovery.

We agree, however, that the original submission did not experimentally isolate the individual contribution of the iterative formulation. We have therefore added a matched ablation, to isolate the iterative formulation, by comparing ABSORB with matched masked-Transformer controls using the same backbone. ABSORB performs better in 64/70 dataset–metric comparisons, including all 14 RMSE comparisons. We have added this ablation study to the Appendix section 2.9, which allows us to sharpen our contribution rather than attributing all gains generically to “diffusion.”

AuthorsOpenReview ↗

Evidence of ablation study on ABSORB vs. Masked-MLP Transformer

Matched ablation: ABSORB vs. masked-Transformer controls trained at fixed mask rates of 0.50 and 0.30 (columns .50 and .30), with the same backbone and tokenization. Held-out recovery; bold marks the best of the three.

Datasetρ↑\rho \uparrow ABSORB / .50 / .30RMSE ↓\downarrow ABSORB / .50 / .30ED ↓\downarrow ABSORB / .50 / .30MMD ↓\downarrow ABSORB / .50 / .30SW ↓\downarrow ABSORB / .50 / .30
Sim D10.922 / 0.907 / 0.9080.609 / 0.675 / 0.667362 / 449 / 4380.144 / 0.125 / 0.11129.7 / 32.7 / 30.5
Sim D20.928 / 0.904 / 0.9100.595 / 0.697 / 0.678268 / 305 / 2970.110 / 0.0997 / 0.092726.4 / 26.5 / 26.7
Sim D30.918 / 0.906 / 0.9100.634 / 0.670 / 0.652220 / 274 / 2810.0826 / 0.0840 / 0.084523.5 / 25.2 / 25.0
Sim D40.942 / 0.917 / 0.9220.549 / 0.670 / 0.644153 / 196 / 2410.0552 / 0.0568 / 0.065320.0 / 22.3 / 24.7
PBMC68k MCAR500.197 / 0.209 / 0.2140.478 / 0.485 / 0.4960.125 / 0.117 / 0.1130.00260 / 0.00199 / 0.001900.0727 / 0.0723 / 0.0671
PBMC68k MNAR250.334 / 0.244 / 0.2340.412 / 0.424 / 0.4840.0517 / 0.0941 / 0.1760.000705 / 0.00169 / 0.003700.0487 / 0.0640 / 0.0931
Bone Marrow MCAR500.392 / 0.324 / 0.3100.703 / 0.834 / 0.7873.37 / 9.45 / 3.900.0209 / 0.0646 / 0.02720.570 / 0.879 / 0.609
Bone Marrow MNAR250.402 / 0.335 / 0.3070.634 / 0.806 / 0.7620.541 / 2.02 / 1.250.00322 / 0.0122 / 0.007560.224 / 0.450 / 0.365
Human Brain MCAR500.280 / 0.239 / 0.2322.543 / 2.638 / 2.731471 / 506 / 5360.0167 / 0.0171 / 0.0175107 / 109 / 110
Human Brain MNAR250.249 / 0.227 / 0.2232.547 / 2.559 / 2.72295.5 / 121 / 2500.00243 / 0.00311 / 0.0072043.5 / 53.3 / 78.5
VISp MCAR500.510 / 0.350 / 0.2972.376 / 2.981 / 2.9281,396 / 2,749 / 3,6780.0147 / 0.0269 / 0.0322153 / 211 / 236
VISp MNAR250.514 / 0.253 / 0.4042.271 / 2.979 / 2.564106 / 535 / 1560.000800 / 0.00451 / 0.0012944.6 / 91.3 / 48.4
Dentate Gyrus MCAR500.481 / 0.464 / 0.4770.709 / 0.949 / 1.09012.8 / 69.3 / 1410.0397 / 0.162 / 0.2600.736 / 2.03 / 3.52
Dentate Gyrus MNAR250.438 / 0.426 / 0.4330.768 / 1.140 / 0.7983.73 / 36.8 / 7.410.0116 / 0.104 / 0.02360.421 / 1.67 / 0.612
ABSORB better13/1414/1413/1411/1413/14
AuthorsOpenReview ↗

Clarifications on selective imputation and biological zeros

Selective imputation and biological zeros…

We agree with the reviewer that selective imputation requires balancing the recovery of genuine dropouts against the preservation of biological zeros. We evaluated different values of the selectivity parameter (α), which determines the expression-frequency threshold used to distinguish potential dropouts from biological zeros within cell clusters. Specifically, the threshold is determined from the α-quantile of observed expression frequencies, allowing more conservative imputation as α increases. The original submission reported the α = 0 operating point to prioritize maximum dropout recovery. However, the sensitivity analysis demonstrates a clear, controllable trade-off, where α = 0.05 substantially reduces erroneous biological-zero imputation while maintaining high genuine-dropout recovery, as shown in the table below.

ABSORB at α = 0 is the original submission; the zero-decision rows use α = 0.02 and α = 0.05.

MethodBiological zeros imputed ↓\downarrowReal dropouts recovered ↑\uparrow
ABSORB, α = 0 (original submission)65.5%91.7%
ABSORB + zero decision, α = 0.0235.9%91.1%
ABSORB + zero decision, α = 0.0512.6%90.2%
SAVER0.1%26.5%
ALRA2.4%72.6%
MAGIC3.4%84.6%
scImpute3.5%81.4%
DCA21.6%83.7%
countsdiff32.8%65.4%
DeepImpute91.3%94.3%

Held-out recovery before → after the zero decision (α = 0 → α = 0.05) on the real datasets. Bold marks the better value; unbolded pairs are ties at the reported precision.

Datasetρ↑\rho \uparrowRMSE ↓\downarrowED ↓\downarrowMMD ↓\downarrowSW ↓\downarrow
PBMC68k MCAR500.197 → 0.1230.478 → 0.4870.125 → 0.1780.00260 → 0.004150.0727 → 0.0868
PBMC68k MNAR250.334 → 0.2820.412 → 0.4230.0517 → 0.05140.000705 → 0.0006710.0487 → 0.0485
Human Brain MCAR500.280 → 0.2982.543 → 2.568471 → 4670.0167 → 0.0166107 → 107
Human Brain MNAR250.249 → 0.2782.547 → 2.56295.5 → 93.10.00243 → 0.0023543.5 → 43.1
Bone Marrow MCAR500.392 → 0.4150.703 → 0.7073.37 → 3.290.0209 → 0.02050.570 → 0.565
Bone Marrow MNAR250.402 → 0.4280.634 → 0.6380.541 → 0.5060.00322 → 0.003010.224 → 0.214
VISp MCAR500.510 → 0.5342.376 → 2.3851,396 → 1,3950.0147 → 0.0146153 → 153
VISp MNAR250.514 → 0.5412.271 → 2.305106 → 1030.000800 → 0.00077244.6 → 44.2
Dentate Gyrus MCAR500.481 → 0.5000.709 → 0.71412.8 → 12.60.0397 → 0.03900.736 → 0.727
Dentate Gyrus MNAR250.438 → 0.4600.768 → 0.7693.73 → 3.590.0116 → 0.01110.421 → 0.410
Zero decision better8/100/109/109/107/10 (2 ties)

Versatility of our proposed algorithm under the selectivity parameter (α) is demonstrated in Figure 16 of Appendix section 2.6.

Importantly, this trade-off is not unique to ABSORB. Methods that preserve biological zeros more strongly, such as ALRA, scImpute, and MAGIC, primarily rely on low-rank approximation, statistical modeling, and graph-based smoothing, respectively, but generally achieve lower dropout recovery. In contrast, deep-learning and generative approaches, including DCA, CountsDiff, and DeepImpute, tend to modify a larger proportion of biological zeros in our evaluation. Nevertheless, at α = 0.05, ABSORB achieves a more favorable balance, retaining higher dropout recovery than the zero-preserving baselines while substantially improving biological-zero preservation. We have included this sensitivity analysis and a discussion of the broader selectivity trade-off in the Appendix section 2.6.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.