- Rating
- 3(+1)
- Confidence
- 3(+1)
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The authors present CRNDiff, a count-native diffusion framework whose forward noising process is defined by stochastic chemical reaction networks, aimed at integer data such as scRNA-seq. With an independent birth-death network, the forward kernel is closed-form (Binomial plus Poisson), enabling exact forward-filtering backward-sampling with a learned posterior and a data-driven terminal noising time based on excess-covariance decay. For conditional generation from a frozen generator, they propose tilted Feynman-Kac steering, which combines per-gene marginal tilting with discriminator-weighted particle correction. Empirical results on synthetic data and the human heart cell atlas suggest higher conditional fidelity than scVI, scANVI, CFGen, and MDLM, especially for rarer cell types, along with preserved marker structure and strong downstream classification utility.
Critical strengths
I personally enjoyed the paper. I find the math very elegant and interesting, especially the use of chemical reaction networks to define a count-native forward process with a closed-form kernel, which makes exact reverse sampling possible. The paper is overall well written: concise but thorough, and the authors are open about the limits of their choices. There are multiple aspects I appreciated, one of which is the attention to devise a forward process that destroys not only single-gene overdispersion but also correlated structure across genes, and the use of this to pick the terminal noising time in a principled way rather than by a validation sweep. Overall, the presented method seems to fit single-cell data very well and in a creative way.
Critical weaknesses
For now, I gave a weak reject based on my questions below, but I am more than happy to reconsider upon rebuttal. I think the main flaw of this submission is that the results section feel a little slim. I am open to increase my score and excited to engage with the authors during the rebuttal phase.
- In my opinion, the motivation presented in the introduction is a bit slim. What is the gap in the literature? What's the problem with latent models with a negative binomial decoder? And with the other discrete approaches?
- Minor, I would add the word "discrete" somewhere in section 2.1's title.
- Minor, L88 define .
- Minor (notation): denotes an earlier forward time in Eqs. (4)–(5) (), but is reused as the reverse-time variable in Eq. (2) (, forward time ), with the opposite meaning. A distinct symbol would avoid confusion.
- I got a bit confused by the notation inconsistency in (9). What's the difference between the matrix and its bold counterpart .
- Please, correct me if I am wrong, but the conditioning pipeline relies on several estimated components. I am wondering about the accuracy of each of them. (i) The marginal tilt uses per-gene count histograms (unit bins over 0–512) from a limited number of target cells (e.g., 3,168 neuronal). Due to overdispersion, these histograms are heavily skewed toward zero with long, sparsely populated tails, so for higher counts, precisely where target-specific expression is expected, the empirical densities are near zero and the ratio is largely determined by sampling noise and by the smoothing and clipping constants. Moreover, the per-target is chosen without a stated procedure. (ii) The FK correction is exact only for the true density ratio (Eq. 23), but it is approximated by the odds of a gradient-boosted classifier on 256 selected genes, used without an upper clip and without a reported calibration check. Given that a conditionally fine-tuned generator outperforms tilted FK on all fidelity metrics (Table 10), the practical benefit of this pipeline over conditional training should be better justified.
- As far as I could understand, the evaluation does not directly assess sparsity, a defining feature of scRNA-seq counts. Table 9 reports only the coefficient of variation of detected-gene counts, not their mean, nor the per-gene or per-cell fraction of zeros. Since the stationary distribution is product-Poisson and the generator is reported as underdispersed (median Fano 2.403 vs. 2.953), comparing zero fractions (e.g., per-gene zero proportion vs. mean expression) between generated and real cells would clarify whether CRNDiff reproduces the sparsity structure of the data.
- I am concerned about possible over-concentration of generated cells for rare targets. For neuronal cells, tilted FK outputs descend from only ~4.5% of initial particles (Anc. 0.045, i.e., ~110 ancestors for 2,500 cells; Table 7). High output uniqueness (0.995) shows that cells are not exact copies, but not that they cover the target's variability, and the reported metrics (per-gene W1, MMD², PCC of means) are not designed to detect concentration. Moreover, UMAP (Fig. 7) does not preserve spread, and the atlas-level PCA (Figs. 6, 8) is dominated by between-cell-type axes. I suggest a PCA fitted on the real target population only, with a comparison of per-PC and total variance between generated and real cells.
- The evaluation on the heart atlas is thorough, combining distributional, marker-level, and downstream metrics with extensive ablations. However, it is limited to a single dataset, with three target populations among eleven classes, the rarest of which (neuronal) still comprises ~2% of cells. It would strengthen the paper to also report results on standard single-cell generation benchmarks, such as those used by CFGen (Palma et al., 2025) and scLDM (Palla et al., 2026), e.g., Dentate Gyrus, Tabula Muris (123 cell types), and HLCA (50 cell types). These would show how tilted FK steering behaves with many, and considerably rarer, cell types and more complex target distributions. The diversity-sensitive metrics reported there (1-NN accuracy, precision/recall) would also help address the concentration concern raised above.
- In Fig. 7d, a substantial fraction of generated neuronal cells appears to lie outside the main neuronal cluster of the test cells (Fig. 7a), notably in a dense group within a different cluster, despite a CellTypist purity of 0.971. This raises the question of whether the marginal tilt steers generation toward neuronal-like per-gene distributions in a region shared with other cell types, rather than into the neuronal population itself. Could the authors show an overlay of real and generated neuronal cells in the same embedding, report which neuronal subtypes (or CellTypist labels) the generated cells are assigned to, and quantify the fraction of generated cells whose nearest real neighbors are neuronal?
Questions for authors
- Might be a naive question, what happens if one sets ? Was the choice dictated by simplicity?
- How should be chosen in practice? Too large a value amplifies noise in the estimated per-gene ratios and degrades fidelity (Table 8, ), while too small a value leaves the FK correction to do all the work, reintroducing the weight concentration the tilt is meant to avoid. The best value between 0.2 and 0.4 is target- and metric-dependent. Could the authors provide a selection procedure or criterion that does not require test-set access, for instance based on ESS or genealogical diversity?
- How do runtime and efficiency compare to continuous approaches?
7 comments
Response to Reviewer z7bw (2/5)
z7bw-W6: Please, correct me if I am wrong, but the conditioning pipeline relies on several estimated components. …
Thank you for this careful analysis of each estimated component. We address them in turn.
1. Noise in the ratio tables (point i). We agree that the high-count tails of the histograms are estimated from few cells. The following two investigations show that this noise has little effect on the generated cells.
First, we rebuilt the neuronal ratio tables from random subsets of the target cells, keeping the generator, discriminator and sampler fixed. Even with 10% of the cells (317 instead of 3,168), the results stay close to those with all cells (Table 14; one frozen generator, one sampling seed, test split): purity 0.969 vs. 0.986, W1 0.060 vs. 0.065, MMD² 0.0119 vs. 0.0131 and coverage 0.812 vs. 0.785.
Second, tilting only the 32 genes with the strongest tilt and leaving the other 1,968 genes untilted gives the same fidelity and coverage as tilting all genes (W1 0.063 vs. 0.063, coverage 0.791 vs. 0.791, purity 0.971 vs. 0.984). The effect of the tilt is therefore concentrated in a few genes, so the noisy tables of the remaining genes contribute little, and the exponent τ < 1 further tempers the ratios. The smoothing constant (ε = 0.5/N) and the clipping range (±log 1000) are now listed in Table 5.
Both investigations are reported in Table 14 (Appendix B.7).
2. Choice of τ (point i). In the revision, we explicitly state a procedure that uses validation cells only (Appendix B.6, Eqs. 45–46). For each candidate τ ∈ {0.2,…,1.0}, we normalize W1, MMD², purity and coverage so that 1 equals the real-cell value, cap the scores at 1, and select the τ closest to the all-ones vector. This gives 0.4, 0.2 and 0.2 for the heart targets and 0.2 and 0.4 for the lung targets. Figure 11 shows how the four scores change with τ.
3. Calibration and clipping of the discriminator (point ii). On held-out cells not used for fitting, the discriminators separate target from proposal cells well and are close to calibrated (Table 11). AUC is the probability that a target cell scores higher than a proposal cell (1: perfect, 0.5: chance); ECE is the expected calibration error after correcting for class proportions (0: perfect); the last column is the share of the total weight carried by the largest 1% of odds. Their definitions are given in Appendix B.6.
| Target | AUC | ECE | Weight carried by the top 1% of odds |
|---|---|---|---|
| Endothelial | 0.931 | 0.060 | 15% |
| Myeloid | 0.988 | 0.028 | 41% |
| Neuronal | 0.992 | 0.011 | 76% |
As the reviewer anticipated, the odds become heavy-tailed for rare targets: for neuronal cells, 1% of the particles carry three quarters of the weight, which explains the genealogical concentration in Table 12. We therefore tested an upper clip at the median odds of held-out real target cells. Its effect is inconsistent across the three generators (test split, unclipped / clipped): W1 0.063 / 0.060, 0.063 / 0.070, 0.070 / 0.063; MMD² 0.0135 / 0.0116, 0.0118 / 0.0167, 0.0167 / 0.0115. We therefore keep the odds unclipped and report this result in Appendix B.6.
4. Benefit over conditional training (point ii). We agree that conditional fine-tuning gives better fidelity when labels are plentiful, and the revised Section 5.2 now states this. Steering a frozen generator is useful in three settings that the revision now evaluates:
-
Few labels (Table 15). With 10% of the neuronal labels (317 cells), fine-tuning yields mostly non-neuronal cells, whereas steering, with ratio tables and discriminator built from the same cells, stays close to its full-label performance: purity 0.959 vs. 0.984 and coverage 0.782 vs. 0.791, whereas fine-tuning reaches 0.359 and 0.256.
-
A target without a training label (Table 16). Steering to UAP1⁺ endothelial cells, defined by a marker gene after training, yields 99.9% UAP1⁺ cells, whereas only 8.6% of endothelial cells express UAP1; label-conditioned generators such as scVI, scANVI, CFGen and CountsDiff could not target this subset without new labels and retraining, whereas CRNDiff reaches it by steering the same frozen generator.
-
Cost (Table 19). Adapting to a new target takes 0.18 h instead of 4.8 h of fine-tuning and leaves the generator unchanged; in exchange, sampling is slower (205 s vs. 31 s per 1,000 cells).
Response to Reviewer z7bw (3/5)
z7bw-W7: As far as I could understand, the evaluation does not directly assess sparsity, a defining feature of scRNA-seq counts. …
Thank you for this suggestion; sparsity is indeed a defining feature of the data, and the original evaluation did not assess it directly. The new Appendix B.8 reports, for all methods and targets, the zero fraction, library size, median Fano factor and variance (Table 18), and the zero fraction against mean expression per gene (Figure 13).
Zero fractions. For endothelial and myeloid cells, CRNDiff reproduces the per-cell zero fraction (0.924 vs. 0.922, 0.927 vs. 0.919), and its gene-wise zero fractions follow those of real cells across expression levels (Figure 13). For neuronal cells, CRNDiff overestimates the overall zero fraction (0.959 vs. 0.949) but, at a given mean, underestimates it for highly expressed genes. Both deviations follow from the count statistics: smaller generated library sizes lower the mean expression of most genes and introduce excess zeros overall, whereas reduced count overdispersion decreases the zero rate at any fixed mean.
Dispersion. The Fano factors quoted by the reviewer (2.403 vs. 2.953) refer to the unconditional pool. For the conditionally generated cells, the median per-gene Fano factor is close to that of real cells (1.59, 1.52, 1.38 vs. 1.53, 1.61, 1.49 for endothelial, myeloid and neuronal cells) and the closest among all methods for endothelial and myeloid cells.
Library size. This analysis revealed a limitation: CRNDiff cells have smaller library sizes than real cells (495 vs. 612, 524 vs. 779, 164 vs. 244). Comparing with cells of the same type in the unconditional samples (Table 17) locates the deficit: for myeloid cells it is already present in the generator, whereas for endothelial and neuronal cells it arises during steering. CountsDiff, which also generates counts directly, shows a similar deficit, while scVI, scANVI and CFGen draw library sizes from real cells and match them by construction. We discuss this limitation in Section 6.
z7bw-W8: I am concerned about possible over-concentration of generated cells for rare targets. …
Thank you for raising this important point and for the concrete suggestion. We followed the suggestion: we analyzed variance of generated cells along a PCA fitted on real target cells (Table 18) and coverage for all methods (Tables 2, 3). The variance ratio of CRNDiff is comparable to that of the best baseline for endothelial and myeloid cells (0.89 vs. 0.92 and 0.81 vs. 0.79, both MDLM) but lower for neuronal cells (0.63 vs. 0.82, CFGen; Table 18). The concentration is real for the rarest target. To locate its source, we compared samplers on the reference generator (neuronal, test split, three sampling seeds; this additional analysis is summarized qualitatively in Appendix B.8):
| Sampler | Variance ratio | Purity |
|---|---|---|
| Tilted FK, τ = 0.2 (selected) | 0.60 | 0.98 |
| FK only (no tilt) | 0.56 | 0.99 |
| Tilt only (no FK) | 0.94 | 0.06 |
| Tilted FK, τ = 1 | 0.99 | lower |
Three observations locate the source. First, the tilt does not cause the concentration: FK steering without the tilt is at least as concentrated as tilted FK steering (variance ratio 0.56 vs. 0.60), whereas the tilt alone keeps the variance close to that of real cells (0.94) but produces mostly non-neuronal cells (purity 0.06). Second, the concentration comes from the FK correction itself: the discriminator odds are heavy-tailed for rare targets (the top 1% of neuronal odds carry 76% of the weight), so the output descends from few initial particles (4.5% for neuronal cells). Using more particles does not remove this concentration. The particles are the candidate cells that the FK sampler propagates and resamples in parallel during generation (5,000 by default; Algorithm 2). On validation cells, four times as many particles more than doubled the number of distinct initial particles from which the output cells descend, but did not increase the variance of the generated cells. The limitation therefore lies in the concentrated weights, not in the number of particles. Third, a larger τ restores variance at the cost of purity and fidelity; at τ = 1 the variance reaches the real value only because library sizes are inflated about fourfold. Appendix B.8 discusses this source, Figure 12 shows the effect of τ on the variance, and the Discussion states the limitation; covering the internal variability of rare targets while keeping high purity is an important direction for future work.
Response to Reviewer z7bw (1/5)
We are very grateful for this thorough and generous review, for the kind words about the mathematics, and for the reviewer's openness to revisiting the score after the rebuttal. The detailed suggestions shaped most of the new analyses in the revision. Following your suggestions, we have substantially extended the results: a second atlas with much rarer targets, a count-native baseline, coverage for all methods, and analyses of each component of the conditioning pipeline. We respond to each point below.
Thank you; we agree that the motivation needed to be more specific. The introduction has been rewritten along these three questions.
Gap: generating rare or newly defined populations without retraining; existing conditional generators need training-time labels.
NB latent models (scVI, scANVI, CFGen): gene dependence passes through a low-dimensional latent; conditioning needs training-time labels.
Other discrete models: categorical diffusion (MDLM) ignores count order; count-native models (CountsDiff and others) condition only on training-time labels. We now compare directly with CountsDiff: CRNDiff is better on 12 of 15 heart target–metric pairs (Table 2) and on both lung targets, where CountsDiff rarely generates the target (Table 3).
The data motivate a count-native model: 94.1% zeros, 56.9% of nonzero entries equal one (Table 4).
Thank you. Section 2.1 is now titled “Continuous-time discrete-state diffusion models”.
Thank you. f is now defined where the generator L is introduced, as a bounded test function f:X→R, so that ∂tE[f(Xt)]=E[(Lf)(Xt)].
Thank you for pointing this out. The reverse-time variable is now u ∈ [0,T] (Eq. 2), so s keeps a single meaning, an earlier forward time with s < t.
Thank you. 𝕊 and S denoted the same object, the stoichiometric matrix whose columns give the change in counts caused by each reaction; the bold version was a typesetting inconsistency. All stoichiometric matrices are now written in blackboard bold, which also distinguishes them from S, the number of species (Section 2.2).