Making Single-Cell Data Distillation Auditable: Traceable Real-Cell Coresets via Discrete Min--Max Selection

Bowen Han; Jianqing Zhu; Lian Zhang; Lingbei Meng; Peize He; Yaodi Luo; Zheng Lu

Making Single-Cell Data Distillation Auditable: Traceable Real-Cell Coresets via Discrete Min--Max Selection

T0 review · 4 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Minmax-CF cuts worst-direction error by about half using only real cells

desk verdict A clearly specified, honest coreset-selection paper whose headline numbers need a sensitivity analysis of its locked candidate cap and per-group budget before I'd fully trust them. read the letter →

arxiv 2607.19426 v2 pith:O6YWNM7I submitted 2026-07-20 q-bio.GN cs.AIcs.LG

Yaodi Luo , Peize He , Lingbei Meng , Bowen Han , Zheng Lu , Jianqing Zhu , Lian Zhang This is my paper

classification q-bio.GNcs.AIcs.LG

keywords single-celldatadistillationtraceablecoresetcharacteristicfunctionmatchingmin-maxoptimizationreal-cellselectionout-of-distributiongeneralizationentropyregularizationauditability

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that single-cell data distillation can be made auditable by forcing the distilled set to consist of original measured cells, and that choosing those cells with an entropy-regularized min-max matching over characteristic-function directions—rather than static matching or geometric coverage—consistently reduces the worst-preserved distribution direction by roughly half across donor, technology, and perturbation shifts. A sympathetic reader would care because training subsets that are traceable to real cells let prediction failures be re-checked against donor, assay, and condition metadata, which synthetic distillation cannot offer. The paper shows that the fidelity gain is direct and reproducible, while downstream utility is largely retained (96.52% balanced-accuracy retention on one annotation benchmark; average parity on a technology-shift benchmark with 2.55× median GPU speedup). It also reports honest failure regimes for rare states, some technology shifts, and unseen perturbation components.

What carries the argument

The load-bearing object is the entropy-regularized discrete min-max problem over a shared pool of characteristic-function directions. An adversary assigns weights to directions proportional to exp(d_j/τ), so poorly preserved projections are upweighted without collapsing onto a single noisy direction; the selector then greedily adds the real candidate cell that minimizes the weighted discrepancy. The greedy loop stays entirely within the empirical support, making the output a subset of observed cells rather than a synthetic profile.

What would settle it

Take a dataset with a known rare population deliberately excluded from the 700-cell candidate pool, run Minmax-CF with the same settings, and check whether worst-direction discrepancy improves while rare-class recall does not; if so, the direct objective is being satisfied without the coverage users actually need.

Watch

Extended reading notes

Core claim

The paper's central claim is that a discrete, entropy-regularized min-max objective over characteristic-function directions—solved greedily within each biological stratum—selects a subset of observed cells that reduces the worst-preserved direction error by roughly half relative to a static matching baseline, across three distinct out-of-distribution regimes, while keeping every selected row a measured cell with original identifiers. The selected submatrix retains original cell IDs and gene symbols, so every row can be traced to measured counts, labels, and provenance; this traceability is the key differentiator from synthetic distillation. The paper shows the fidelity gain is direct and rep

Load-bearing premise

The whole construction relies on a fixed candidate pool of at most 700 training cells per biological stratum and a budget of at most 50 cells per group; if a rare or condition-relevant state is absent from that cap or cannot be reached within 50 cells, no reweighting can preserve it.

Editorial extensions

If this is right

If the central claim holds, any single-cell training object can be compressed to a few hundred real cells per stratum while retaining traceability and most of the full-data utility, enabling cheap repeated retraining and audits.
Worst-direction discrepancy is a reproducible, direct-objective gain that does not by itself guarantee downstream accuracy; fidelity and utility must be tracked separately.
The retained-ID contract makes failure regimes inspectable: a rare-state miss, a technology shift, or an unseen perturbation component can be traced to the exact cells, donors, and conditions present in the training support.
The method introduces no inference-time selector or generator, so the distilled object is an ordinary sparse slice of the original count matrix and can be reused across models and hyperparameter searches.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

An obvious extension the paper does not pursue is making the candidate cap and per-stratum budget adaptive to rarity: a sensitivity study over cap and budget would reveal whether the rare-state recall shortfall is a coverage artifact or an algorithmic limit.
Because selection is per-stratum and additive, the same min-max principle could be applied hierarchically—first choosing strata, then cells within strata—to handle very large atlases with many strata without exceeding the per-stratum cap.
The fact that geometric coverage methods win on local-distance metrics while Minmax-CF wins on projection-fidelity metrics suggests a hybrid objective combining adversarial projection matching with local geometric coverage might dominate both in downstream utility; the paper itself leaves this as a possibility.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The short version: this is a genuinely new formulation and a carefully honest evaluation. The authors define traceable single-cell distillation as retaining original cell IDs and gene symbols, and they propose Minmax-CF, a discrete entropy-regularized adversarial characteristic-function selector. That combination is not in the cited literature, and the paper does not oversell it. They consistently separate direct fidelity from downstream utility, report failure regimes plainly, and use leakage-resistant donor/technology/perturbation splits. The 42–52% worst-direction CF reductions and the 96.52% MS BA retention are consistent with the stated objective; the downstream gains are modest and presented as such. Also, the auditability argument is real: if you keep measured cell IDs, you can actually go back and inspect what the model saw.

The soft spots are real but not fatal. The biggest is exactly what your stress-test note flags: the method depends on a locked candidate pool of 700 cells per stratum and a budget of at most 50 cells per group. The paper gives no sensitivity analysis for either. The rare-state recall failure on MS (0.704 vs the 0.90 target) and the near-zero Norman seen-0/unseen-single correlations are consistent with coverage limits imposed by that cap and budget. It could be that a different cap or budget changes both the gains and the failures, and we just don't know. That is a load-bearing gap, not a minor omission. Second, there is no code or data release, so the numbers are not independently reproducible. Third, some hPancreas comparisons use unequal seed counts (Random and Fixed-CF with n=9 vs PCA and Minmax-CF with n=27); the authors disclose this, but it does weaken a few of the head-to-head claims. The fidelity–utility coupling is weak (correlations near zero or negative), which the paper acknowledges and treats as a feature of the analysis rather than a bug.

Who gets value: anyone working on dataset condensation, single-cell data management, or reproducibility tooling. The traceability framing is worth taking seriously even if the specific optimizer is not the final word. It deserves a serious referee. My recommendation: send it out, but require a sensitivity analysis for M, tau, the candidate cap, and the per-group budget, plus code and data for the main experiments. Without those, the central empirical claims remain conditional.

Referee Report

4 major / 4 minor

Summary. The paper formulates traceable single-cell data distillation as selecting a subset of original cell IDs and gene symbols under fixed budgets, and proposes two real-cell selectors: Fixed-CF, which greedily matches a static set of characteristic-function (CF) directions, and Minmax-CF, which solves an entropy-regularized discrete min–max problem that adaptively upweights currently poorly preserved CF directions. The evaluation covers three datasets with donor-, technology-, and perturbation-level held-out shifts, and measures distribution fidelity, downstream utility, and compute cost separately. The reported headline results are: Minmax-CF reduces soft/worst-direction CF discrepancy by roughly 42–52% relative to Fixed-CF; it retains 96.52% of Full balanced accuracy on MS; on hPancreas it approximately matches Full with a median 2.55× GPU speedup; and on Norman it achieves the lowest pathway MAE among compressed methods. The paper is transparent about failure regimes: rare states, some technology shifts, unseen perturbation components, and weak fidelity–utility coupling. It also clearly separates direct-objective gains from downstream superiority claims.

Significance. If the results hold, the paper makes a useful contribution by reframing dataset distillation for single-cell data under a traceability constraint: every selected row is an assayed cell with recoverable provenance, which enables audit of unexpected predictions. The leakage-resistant evaluation design (preprocessing and selection fit inside training folds; donors/technologies/perturbation groups held out) is a strength, as is the explicit separation of fidelity, utility, and cost. The method is also computationally modest and, because it outputs a sparse slice of the original matrix, avoids the infrastructure burden of generative distillation. The paper is appropriately cautious: it does not claim universal downstream superiority, and it identifies regimes where simple baselines match or exceed Minmax-CF. However, two issues currently limit the strength of the central claims: (i) a numerical inconsistency in the reported pathway-MAE improvements, and (ii) the absence of sensitivity analysis for the locked candidate-pool cap and per-stratum budget, which are load-bearing for the claimed consistency of the fidelity gains and for the acknowledged rare-state failures.

major comments (4)

[§Results / Figure 3 / Figure 8 / Table 2] There is a direct numerical inconsistency in the pathway-MAE claims. The text states that Minmax-CF reduces pathway-mean MAE by 10.8%, 8.7%, and 10.6% relative to Fixed-CF, and Figure 8 shows Minmax-CF at 89–91% of Fixed-CF for all three datasets. But Table 2 reports Norman Fixed-CF path MAE = 0.03516 and Minmax-CF = 0.03465, a reduction of only 1.45%; Random (0.03530) is not 140% of Fixed-CF as Figure 8 implies. This discrepancy is not a minor typo: the claim 'lowest pathway error among compressed methods on Norman' rests on a tiny margin (0.0005) that is visually magnified in Figure 8. The authors need to reconcile the reported percentages with the raw values and state which gene setting and normalization the figures use.
[§Locked configuration and cost / §Results] The locked configuration (M=512, τ=0.05, candidate cap 700 cells per stratum, budget ≤50 cells per biological group) is load-bearing for the headline 42–52% worst-direction discrepancy reductions. The greedy selector can only add cells from the capped pool, so if a condition-relevant state is absent from those 700 candidates or cannot be represented within 50 cells, no reweighting can preserve it. The paper's own results—MS rare recall 0.704 vs. the 0.90 target, and Norman seen-0/unseen-single correlations near −0.03—are consistent with this coverage limit. No sensitivity analysis is reported for either the candidate cap or the per-stratum budget. The authors should show how the fidelity gains, rare-state recall, and downstream utility vary when, e.g., the cap is 350 or 1400 and the budget is 25 or 100, at least on one dataset, to establish that the reported properties are not artifacts
[§Eq. (1) / §Figure 3] The soft and worst-direction CF discrepancy reductions in Figure 3 are computed on the same 512-direction pool Ω that Minmax-CF optimizes in Eq. (1). The greedy procedure is explicitly selecting cells to minimize the discrepancy on these directions, so the reductions are partly in-sample for the selector objective: they demonstrate that the optimizer works, but do not by themselves establish that Minmax-CF improves worst-direction fidelity in a direction-agnostic sense. The paper acknowledges this in prose ('lower worst-direction discrepancy validates the declared fidelity objective'), but the abstract and contributions present the reductions as a headline finding. Please add an explicit statement, or better an evaluation, using a disjoint held-out set of CF directions (e.g., a different random seed for Ω) to show that the reductions generalize to directions not used during selection.
[§MS donor-OOD annotation / Table 2] On MS, Minmax-CF's rare-class recall (0.704) is below the prespecified 0.90 target and below Fixed-CF's recall (0.721), even though Minmax-CF improves the CF discrepancy. Since rare-state preservation is a stated motivation for traceable real-cell coresets, this trade-off deserves a dedicated discussion: why does the adversarially reweighted selector sacrifice rare recall relative to Fixed-CF, and could the objective in Eq. (1) or the per-stratum budget be adjusted to improve rare-state retention without losing the discrepancy gains? Currently the paper reports the gap but does not analyze its cause beyond the qualitative coverage concern.

minor comments (4)

[§hPancreas technology-OOD annotation] The technology-cluster bootstrap interval [−0.0411, 0.0455] for mean paired ΔBA is based on only nine technologies. Please state explicitly that this interval should be interpreted with caution given the small number of clusters, and perhaps report the distribution of fold-level means as well as the 27 fold-seed units.
[§Norman perturbation-support analysis] The seen-2/seen-1 correlation differences (Minmax-CF 0.833 vs. 0.537) are reported without uncertainty intervals. Given only three seeds, a bootstrap over perturbation groups would strengthen the claim that support visibility drives the degradation.
[§Figure 8] The heatmap legend states 'darker blue means lower error', but the exact values are printed in every cell. Consider using a consistent color scale across datasets, since the current normalization to Fixed-CF makes the Random/PCA columns visually similar to the MS column despite different absolute scales.
[§Introduction / Related Work] The formulation-level comparison with scDD is appropriate, but the paper would benefit from one sentence on whether scDD or other synthetic methods could also be made traceable by outputting nearest real-cell IDs, to clarify the specific advantage of the real-cell constraint.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper correctly labels CF-discrepancy reductions as direct-objective evidence and supports the central claim with independent OOD utility/cost analyses.

full rationale

Minmax-CF's soft and worst-direction CF reductions are aligned with the entropy-regularized min-max objective in Eq. (1), but the paper does not present them as independent predictions; it explicitly states that 'Lower worst-direction discrepancy therefore validates the declared fidelity objective, not universal downstream superiority.' The CF evaluation uses the same held-out direction pool as selection, yet this is acknowledged as direct-objective evidence rather than an external benchmark. The independent evidence is the downstream OOD utility and cost analysis, in which selection never observes test labels and which reports weak or negative fidelity-utility correlations (Pearson r = 0.29 and -0.17 in Figure 5), so those results are not forced by the selection objective. No load-bearing self-citation is present: the cited min-max CF work (Wang et al. 2025) has no overlapping authors, and the core derivation (Eq. 1-2 plus the discrete greedy algorithm) is provided in the paper. The locked candidate cap and per-stratum budget are robustness/sensitivity concerns rather than circularity, and the paper is transparent about its failure regimes, including rare-state recall below target and support-dependent Norman correlations. Therefore the derivation chain is self-contained and no circular step meets the quoted-evidence threshold.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical entities. It relies on four hand-chosen hyperparameters, a fidelity metric, a stratification assumption, a greedy heuristic, and a train-only leakage boundary. The central claim rests primarily on the domain assumption that characteristic-function discrepancy is a meaningful objective for real-cell coreset selection.

free parameters (4)

M (number of CF directions) = 512
Chosen by hand; all evaluations use a shared 512-direction pool. No sensitivity analysis is reported.
τ (entropy regularization temperature) = 0.05
Controls how concentrated the adversary's weights become; fixed before downstream evaluation without sensitivity analysis.
candidate cap per stratum = 700 cells
Caps the greedy search pool per biological stratum; rare states outside the cap cannot be selected.
budget per biological group = ≤50 cells
Maximum number of selected cells per stratum; combined with the cap, this defines the compression ratio.

assumptions (4)

domain assumption The empirical characteristic function over a finite pool of 512 directions is a valid fidelity measure for single-cell expression distributions.
Used to define d_j(S) in Eq.1; if this metric does not capture biologically relevant variation, the selector's objective is misaligned with downstream utility. The paper itself shows weak fidelity–utility coupling.
domain assumption Biological strata c_i are known and used to stratify selection, and within-stratum diversity is captured by the candidate cap.
The method concatenates per-stratum selections; if strata are misspecified or a rare state falls outside the cap, it cannot be preserved.
ad hoc to paper The greedy herding procedure produces a good approximate solution to the combinatorial min–max problem in Eq.1.
No convergence or optimality proof is given; support is empirical traces in Figure 2.
domain assumption Preprocessing and frequency pools fit on training splits only are representative of held-out conditions.
The leakage-resistant protocol assumes that train-only fitted statistics transfer; some technology-fold failures suggest this can break.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Making Single-Cell Data Distillation Auditable: Traceable Real-Cell Coresets via Discrete Min--Max Selection." pith.science (2026). https://pith.science/paper/O6YWNM7I

@misc{pith2026260719426,
  author       = {Pith},
  title        = {Pith review of: Making Single-Cell Data Distillation Auditable: Traceable Real-Cell Coresets via Discrete Min--Max Selection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O6YWNM7I}},
  note         = {Machine review of arXiv:2607.19426}
}

read the original abstract

Large single-cell datasets are expensive to store, curate, and repeatedly reuse for model training. Data distillation can reduce this burden by building smaller training sets. However, many existing methods rely on synthetic cells. These synthetic cells do not retain direct correspondence with assayed cells and genes. This limits source-level inspection and biological traceability. Moreover, real-cell expression matrices are often sparse and noisy. In light of these challenges, we propose Minmax-CF, a label-aware characteristic-function selector for traceable single-cell data distillation. Minmax-CF formulates compression as a discrete min--max selection problem over characteristic-function directions. It uses entropy-regularized maximization to emphasize the least preserved directions. Greedy minimization ranks cells and genes by how much they reduce the resulting weighted error. The method alternates cell and gene selection under explicit axis-specific budgets. Across five coarse-lineage benchmarks and five compression budgets, Minmax-CF retains 95.3% of the Full-reference macro-F1 on average, with gaps that exceed one per-seed standard deviation. It also retains exact source-cell indices and original gene symbols. Compared with size-matched synthetic PCA-Centroid and Distribution Matching (DM) baselines, Minmax-CF achieves higher coarse-lineage macro-F1 in 24 of 25 comparisons against each baseline. It exceeds their average performance by 10.4% and 17.4%, respectively. Retained cells can also be projected onto independently computed embeddings for direct biological interpretation.