acceptodds
Withdrawn from ICLR 2027

Geometry-Constrained Residual Projection with Optimal Transport for scRNA Cell-State Analysis

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Observed zeros in single-cell RNA sequencing (scRNA-seq) data couple two problems that are usually treated separately: expression imputation and cell-state progression inference. Imputation seeks to recover expression missed by the assay, whereas optimal transport (OT) infers progression from distributions of cell states. However, imputation changes the same cell-state geometry from which OT infers progression. To limit this change, we introduce Geometry-Constrained Residual Projection (GCRP), which treats existing imputers as generators of candidate residuals. GCRP accepts only corrections that are supported by neighboring cells, have a controlled amplitude, and satisfy cell-displacement and population-level OT limits. Per-cell displacement bounds induce Wasserstein bounds on population displacement and on the distortion of pairwise group distances. Candidate selection relies only on agreement with observed positive expression and on OT over clusters of raw expression. GCRP itself uses no biological labels or external measurements, which serve for evaluation. Across controlled-corruption, CITE-seq, disease-progression, and time-course experiments, GCRP and its support gate keep the population geometry used for progression inference. Under controlled corruption, full graph imputation triples the error of the relative state geometry against an uncorrupted reference, whereas GCRP keeps this error at the level of raw data and lowers held-out expression RMSE by 5%. Applied on top of each of five published imputers, GCRP lowers that imputer's distortion of the population geometry 3.7- to 8.2-fold. On CITE-seq cells, scVI and SAVER lower the correlation between RNA and surface-protein levels, and GCRP applied on top of them raises it in every seed, from 0.146 to 0.290 for scVI. At the default radius, the Wasserstein bounds certify that the displacement projection preserves 45% of the raw orderings of group distances. These results show that progression geometry can set how much uncertain expression correction is admitted.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
scRNA-seq imputation, optimal transport, geometric constraints, cell-state progression, trajectory inference, Wasserstein bounds

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_TkV7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Geometry-Constrained Residual Projection, or GCRP, a framework that constrains scRNA-seq imputation to preserve cell-state geometry used in optimal transport-based progression analysis. Rather than developing a new imputation model, GCRP treats existing imputers as sources of candidate residuals and selectively accepts corrections based on neighborhood support, correction magnitude, cell-level displacement bounds, and population-level OT constraints. The paper derives Wasserstein-based bounds on population displacement and relative distance ordering. Experiments on CITE-seq datasets, controlled corruption, disease progression, and time-course data show that GCRP reduces geometric distortion compared with several existing imputers while retaining some expression recovery.

Critical strengths

GCRP separates candidate expression recovery from the decision of whether a correction should be accepted. By combining neighborhood support, cell-level displacement constraints, and population-level OT checks, the method can be applied to different existing imputers without retraining them. Propositions 1–3 further establish how bounded cell displacement limits changes in population Wasserstein distances and preserves sufficiently separated group-distance orderings relative to the observed geometry.

Critical weaknesses
  • Propositions 2–3 establish bounds on pairwise Wasserstein distances and their relative orderings between cell populations, but these guarantees do not directly address the stability of OT couplings or inferred cell-state trajectories. In Appendix Table 22, GCRP only reduces the held-out zebrafish interpolation error from 0.887 to 0.886, while a simpler neighbor-mean correction achieves 0.882. It remains unclear whether preserving population-level distance geometry actually leads to more reliable progression inference.
  • Propositions 1–3 establish geometric stability guarantees for fixed affine reference maps and unregularized Wasserstein distances, while the practical implementation also uses nonlinear reference maps and Sinkhorn costs computed in separately fitted embeddings. Appendix B acknowledges that the nonlinear displacement constraints are not guaranteed to hold under the proposed scaling procedure. It remains unclear to what extent the theoretical guarantees apply to the geometric constraints enforced by the actual implementation.
Questions for authors
  • Could the authors provide stronger evidence that preserving population-level Wasserstein distances and their orderings improves the stability or accuracy of downstream progression inference, particularly compared with the simpler corrections evaluated in Appendix Table 22?
  • Could the authors clarify which theoretical guarantees remain applicable to the nonlinear reference maps and Sinkhorn-based geometry checks used in practice? How do they verify whether the implemented algorithm satisfies the nonlinear displacement constraints?
Reviewer_6dg1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Geometry-Constrained Residual Projection (GCRP), a postprocessing layer for existing scRNA-seq imputers. It treats their outputs as residual proposals and accepts nonnegative corrections on observed zeros using expression support, amplitude controls, cell-displacement limits and population-level OT checks. The full selector combines proposals into residual variants and strength candidates, then maximizes a heuristic score among candidates that pass the geometry checks, with fallback shrinkage when necessary.

For fixed affine reference maps, the analysis relates cell-displacement budgets to Wasserstein perturbation bounds and preservation of sufficiently separated group-distance orderings. Experiments include CITE-seq panels, controlled corruption, disease-state and time-course analyses. The reported results show reduced distortion relative to raw geometry and some improvements against independent references, while also revealing a recovery–preservation tradeoff.

Critical strengths
  1. An interesting use of inter-cluster OT distances. The paper considers OT distances between cell clusters and compares these distances before and after imputation to assess their stability. I consider this particular perspective to have some novelty and to provide a useful way to examine how imputation affects relationships between cell populations.

  2. A coherent conditional rationale. When the origin of an observed zero is uncertain, limiting unsupported changes is a defensible conservative objective. The fixed affine reference construction gives an explicit displacement certificate, and the natural cellwise coupling and Wasserstein triangle inequality translate this into interpretable population and ordering bounds (Section 4.1; Appendix A). These arguments explain the intended stability mechanism while keeping its assumptions visible.

Critical weaknesses
  1. The methodological novelty needs sharper positioning. Concerns about imputation-induced false signals predate this paper (Andrews and Hemberg, 2018), and conservative zero-preserving imputation is already an explicit goal of ALRA (Linderman et al., 2022). CROT (Liu et al., 2026) also employs optimal transport with cluster-level regularization for single-cell imputation. I nevertheless find the paper’s specific comparison of inter-cluster OT distances before and after imputation to be a novel perspective. Its explicit geometry budget also provides a distinct way to control residual proposals from existing imputers. However, the paper should more clearly establish what methodological advance this design offers beyond prior OT-based imputation and simpler conservative corrections. In my view, the novelty lies mainly in this particular perspective and formulation, while the incremental methodological contribution appears limited. The experimental results in Tables 19 and 21 further show a tradeoff rather than a consistent recovery advantage over the evaluated baselines, leaving the added practical value of the full selector insufficiently established.

  2. The rationale and guarantees remain conditional on the raw anchor and representation. A reliable observed geometry makes conservative correction sensible, but missing capture can already distort that anchor, as Appendix H acknowledges. Local support and a large heuristic score do not by themselves establish correct recovery. The formal bounds concern unregularized Wasserstein distances in fixed affine maps, whereas the operational checks use Sinkhorn costs in refitted embeddings; the nonlinear maps are also heuristic limits. These distinctions do not invalidate the affine proofs, but they restrict what those proofs establish about the full selector and downstream progression. The method should be presented as controlling declared geometric change, with recovery benefits requiring separate evidence.

Questions for authors
  1. Which part of the acceptance framework constitutes the main methodological advance over existing conservative imputation and simpler residual controls? Please explain why the combination of entry support, displacement limits and population-level OT screening is needed for the stated goal, while distinguishing this contribution from the standard ingredients of the stability proofs.

  2. Can the authors provide one algorithm defining candidate construction, proposal weighting, residual variants, strength choices and fallback operations, and state which guarantees apply to each step? Please distinguish directional displacement scaling and maximization of L, fixed affine Wasserstein guarantees and operational Sinkhorn checks, including the status of the nonlinear maps and small-fill branch.

  3. Under what conditions is raw geometry a defensible anchor for the intended downstream problem, and how should users interpret the output when those conditions are doubtful? In particular, agreement among imputers is not independent biological evidence, so what observable information supports accepting or rejecting a correction in that situation?

  4. For the experimental design, can the authors consolidate neighborhood spaces, alpha settings, cluster construction, internal acceptance rules, external evaluation thresholds and seed-dependent steps into one protocol table? Please distinguish RNA-only selection from annotation-assisted sampling and evaluation, and explain what the same-panel default tuning and repeated seeds do and do not establish.

  5. For the experimental results, which independent-reference benefit most clearly justifies the full selector over simpler conservative controls? Building on Tables 17–19, 21 and 24, please clarify the recovery–geometry tradeoff and its dependence on corruption or anchor quality under a common comparison protocol. This would help identify when GCRP adds practical value despite the stronger recovery RMSE of some simpler controls.

Reviewer_1bAEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that single-cell RNA imputation can improve entry-wise reconstruction yet distort the cell-state geometry used in optimal-transport (OT) progression analysis. It proposes GCRP as a model-agnostic postprocessor: existing imputers supply residual proposals, a local-support gate restricts which zeros can change, per-cell scaling bounds displacement in reference embeddings, and a population-level OT selector rejects proposals that change raw-expression group distances beyond a prescribed budget. Elementary coupling and triangle-inequality results give Wasserstein stability guarantees relative to raw expression under fixed affine maps. Across five base imputers on three CITE-seq panels, the reported OT error decreases by factors of 3.7–8.2. Additional tests cover synthetic dropout, RNA–protein concordance, clinical states, and time-course interpolation.

Critical strengths
  1. A meaningful problem distinction. Recovering missing expression and preserving biologically important relationships are different objectives. The paper articulates the risk of optimizing one at the expense of the other, consistent with earlier warnings about imputation artifacts.
  2. Broad compatibility. The proposal/acceptance separation can be placed on top of five heterogeneous imputers (scVI, MAGIC, ALRA, DCA, SAVER), offering practical value independent of a particular generative model (Table 1).
  3. Transparent and correct scoped stability arguments. The paper carefully states that the Wasserstein bounds apply to fixed reference maps and to distortion relative to raw data, not to unknown true biology (Propositions 1–3; Appendix A). This is a useful formal guarantee even if its ingredients are standard.
Critical weaknesses
  1. The headline OT improvement is partly enforced by construction. GCRP directly selects candidates based on closeness to raw-cluster OT geometry, then reports reduced OT error against raw geometry. This establishes compliance with the chosen budget, but is not independent proof of better inferred progression. Main comparisons should foreground held-out ground truth and Pareto curves at matched recovery, filled-zero fraction, and geometric distortion.
  2. The guarantee does not establish recovery of true cell-state geometry. If dropout has already corrupted the raw manifold, preventing movement away from it may retain its errors. Moreover, the rigorous bounds concern unregularized Wasserstein distance in fixed affine maps, whereas some evaluations use Sinkhorn discrepancies in PCA spaces refitted for each output; nonlinear maps are only heuristically constrained. The paper discloses these limits, but the 'certified 45%' refers only to raw distance orderings, not to biological progression accuracy.
  3. Methodological novelty is primarily in integration and selection. Support thresholds, radial clipping, OT discrepancy constraints, and the coupling/triangle-inequality bounds are established tools. The valuable new proposal is to combine them as a residual acceptance policy for this use case. Stronger matched-budget baselines and a clearer comparison with simpler support-gated shrinkage are necessary to demonstrate which part produces nontrivial benefits.
Questions for authors

See the weakness.

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

Related papers

Discussion (0)

Sign in to comment.