acceptodds
Withdrawn from ICLR 2027

SelCell: Self-Certifying Cell Velocity from Destructive Snapshots

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Spatial transcriptomics destroys the tissue it measures, so every cell is observed exactly once: cell velocity—where cells are going—must be inferred from a single snapshot and can never be checked against a later observation. Current practice reads RNA velocity as cell motion, an implicit assumption that real data do not always support: coherent arrows are drawn even on embryonic tissue where the motion signal is indistinguishable from a shuffled control, and existing confidence scores do not separate a method’s reliable cells from the rest. We replace the assumption with an explicit observation model that separates a cell’s intrinsic expression change from the change caused by its motion across the tissue, and invert it in closed form—training-free, compatible with any RNA-velocity backbone. Three per-cell certificates then follow analytically, each answering a different question: where to trust the recovered arrows, how much (direction-wise error bars, order-optimal for the noise they certify), and what is being measured (motion, or a wave of expression sweeping past still cells—bounded by planting synthetic waves). On real tissue the error bars are calibrated and remain valid on the hardest cells where split conformal bands collapse; ranking cells by the validity certificate is the only ranking we found that improves the kept cells. On certified tissue SELCELL outperforms the strongest baseline (CBDir 0.232 vs 0.183, non-overlapping intervals) while abstaining where no motion signal exists. Code, raw-read rebuild pipelines, and the preregistered, hash-sealed protocols are in the supplementary material. We believe this recipe—replace an unchecked assumption with an explicit observation model, then certify what its inversion returns—extends to any estimand that can never be observed twice.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
spatial transcriptomics, RNA velocity, cell velocity, uncertainty quantification, selective prediction, calibration, identifiability

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_i69WOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors introduce SelCell a novel method for calculating cell velocity from spatial transcriptomic snapshots. The authors propose to replace the implicit expression changes and physical motions with an explicit model. Under their assumptions, they can derive 3 certificates about the resulting velocities trustworthiness and the measurements. The authors evaluate SelCell on two synthetic benchmarks and 5 real transcriptomic snapshots against established velocity estimation methods.

Critical strengths

None

Critical weaknesses

The empirical evidence is limited to small datasets.

Can the authors please add a simple example with synthetic data to illustrate their mechanism better.

No discussion about the validity and assumptions behind the observation model. Also while the proposed mechanisms for RNA readouts are sensible they are unlikely to reflect the whole biology and is neglecting important components such as some cells changing expression in response to cytokines, the differentiation of tissues and responses to external stimuli e.g. injury or inflammation. Hence, the certificate cannot provide explicit guarantees if these assumptions break which is not discussed much.

The article is not written clearly, several sections appear to be generated with extensive LLM usage and it the certification parts are repeated several times.

Questions for authors

I disagree with the notion that cell velocity must be observed from a single snapshot. While omics approaches are indeed destructive measurements there are plenty of methods to measure temporal samples from similar samples and align the snapshots using optimal transport and learn temporal models such as MOSCOT, CellBRIDGE, NicheFLow etc.

Can the authors please show an example of their learned velocities against gound truth that is easier to parse than Figure 7 and include it in the main text. I am also struggling to see in that image why selcell is performing better as both fields look quite different from ground truth.

The last sentence of the abstract is too strong of a claim for the evidence presented.

Why is the image in Fig 1 not a spatial transcriptomic image?

Limitation section does not discuss that the problem is fundamentally ill posed, it is impossible by definition to learn a velocity from a single snapshot, one can only estimate certain processes under assumptions about the cell expression and movement that can be biased.

Reviewer_HfTMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SelCell, a training-free framework for estimating cell velocity from spatial transcriptomic snapshots. The method introduces an observation model that separates motion-driven expression changes from intrinsic expression dynamics and derives a closed-form velocity estimator. Three certificates are developed to assess the reliability of recovered velocities, quantify direction-wise uncertainty, and distinguish physical motion from coherent expression waves. Experiments on synthetic benchmarks and spatial transcriptomic datasets evaluate velocity estimation, uncertainty calibration, and selective abstention. SelCell improves directional alignment over the strongest baseline on the primary thymus section and abstains when the observed signal cannot be distinguished from shuffled controls.

Critical strengths

The paper introduces an explicit observation model that distinguishes intrinsic expression dynamics from physical cell motion and analyzes their inherent non-identifiability from a single snapshot. Based on this formulation, the authors derive a closed-form estimator for the identifiable velocity component and develop three certificates to determine where the estimates can be trusted, quantify direction-wise estimation errors, and assess whether the observed signal may arise from expression waves rather than physical motion. This formulation provides reliability and uncertainty information alongside velocity estimates instead of directly interpreting RNA velocity as cell migration.

Critical weaknesses
  • Certificate 3 derives an upper bound on expression waves in thymus S1 using synthetic-wave injections. However, this result relies on predefined gene partitions and assumptions about subset residuals, while the injection experiments only cover a restricted family of tissue-coherent waves. It remains unclear whether this bound can reliably exclude real confounding waves, particularly when actual expression dynamics differ from the injected patterns or the gene-partition assumptions are violated.
  • Section 4.4 reports that on thymus S3, cells with the highest R² values exhibit worse directional alignment, even though Certificate 1 passes at the tissue level, as shown in Appendix Table 3. Although the authors attribute this failure to structured residuals, the proposed reliability criterion relies on assumptions that may not hold in real tissues. It remains unclear how such violations can be reliably detected before using R² to select trustworthy cells. This raises concerns about whether Certificate 1 can consistently identify reliable velocity estimates under realistic tissue conditions.
Questions for authors
  • Certificate 3 relies on synthetic-wave injections to bound potential confounding. Beyond the gene-partition contamination experiments in Appendix B.3, could the authors clarify how reliable this bound remains when real expression dynamics differ from the injected wave patterns or violate the subset-residual assumptions?

  • On thymus S3, Certificate 1 passes the tissue-level test, but cells with higher R² exhibit worse directional alignment. Could the authors clarify how such structured-residual failures can be detected using observable information before applying R²-based selection?

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

This paper proposes SelCell, a framework for inferring cell velocity from RNA velocities estimated by existing single-cell or spatial transcriptomics methods. It models RNA velocity as a combination of intrinsic expression dynamics and expression changes associated with movement through space. The method estimates local spatial expression gradients and uses linear regression to project RNA velocity onto the subspace spanned by those gradients.

The authors introduce three “certificates”: a goodness-of-fit statistic, model-derived directional error bars and confidence intervals, and a biological gene-subset negative control which is assessed by injecting synthetic wave-like perturbations into RNA velocities These quantities govern blending, uncertainty reporting, and whether the output is interpreted as physical motion.

Critical strengths
  1. The paper addresses an interesting and important question: whether spatial transcriptomics and RNA velocity can distinguish physical cell movement from changes in gene-expression programs.

  2. The estimator is simple, interpretable, and compatible with existing RNA-velocity methods.

Critical weaknesses
  1. Strong biological assumptions: Projecting RNA velocity onto spatial expression gradients establishes alignment, but does not establish movement. The assumptions supporting that interpretation need stronger validation.

  2. Limited real-data evidence: The only positive benchmark is one thymus section, with two additional sections from the same donor.

  3. Model-dependent certification: The certificates largely assess consistency with the proposed model. Confidence intervals and detection of model-generated perturbations provide limited independent evidence for the biological interpretation.

  4. Unclear practical implications. Although certificates govern blending and reporting, the biological conclusions users may draw from passing, failing, or mixed outcomes remain insufficiently clear.

Questions for authors
  1. What testable assumptions permit interpreting w−cw-c as physical motion, and how should users verify them?

  2. What false-negative or coverage guarantee supports converting detector silence and synthetic-injection sensitivity into a bound on natural confounding?

  3. What downstream conclusions are justified for each certificate outcome, particularly when the method falls back to the backbone prediction?

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

Related papers

Discussion (0)

Sign in to comment.