acceptodds
Under review as a conference paper at ICLR 2027

Learning Early Treatment Strategy from Snapshots for mRNA-protein Regulatory Networks

Abstract

Early disease detection with snapshot data has been effectively addressed by the Dynamical Network Biomarkers (DNBs) theory. After early disease detection, it is crucial to consider early medical treatment to prevent it. This paper presents a novel framework for identifying mRNA-protein regulatory systems from snapshot data and designing interventions. We estimate the state covariance of mRNA-protein expression using multi-episode snapshot samples and identify the underlying continuous-time dynamics by solving a Lyapunov-based regression problem. We provide finite-sample guarantees on the estimation accuracy of the system matrix and its dominant eigenvectors, which are essential for downstream treatment design. Building on these estimates, we formulate an optimal re-stabilization strategy that minimizes input energy with desired spectral shifts. To ensure practical feasibility, we further propose a diagonal re-stabilization scheme that identifies key regulatory nodes using a first-order eigenvalue sensitivity analysis. Validations on synthetic mRNA-protein network and real dataset demonstrate that our method accurately identifies regulatory nodes under snapshot data conditions and significantly outperforms existing baselines.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
system identification, dynamical systems, snapshot data, single-cell RNA-seq, gene regulatory networks, master regulator identification, structured identifiability, finite-sample guarantees, Lyapunov equation, spectral perturbation, critical transitions, high-dimensional low-sample-size

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

This paper investigates how to identify mRNA-protein regulatory network dynamics from multi-episode single-cell snapshot data and design early intervention strategies. By leveraging structural zero constraints of the two-modality regulatory network, the authors recast continuous-time system identification as an inverse Lyapunov linear regression problem and establish finite-sample error bounds for the system matrix and dominant spectral projector via cross-episode kernel conditional density estimation. Building on these estimates, the framework ranks candidate regulatory nodes using first-order eigenvalue sensitivity (participation factors) to achieve spectral re-stabilization, validating master-regulator identification on a 90-dimensional synthetic modular network and a 50-dimensional single-cell hematopoietic differentiation dataset.

Critical strengths
  1. This paper incorporates structural sparsity priors into the inverse Lyapunov equation, establishing a complete theoretical pipeline with finite-sample error propagation bounds from kernel covariance estimation to the system matrix and its dominant spectral projector (Proposition 1 and Theorems 1–3) without requiring explicit trajectory reconstruction.

  2. By leveraging first-order eigenvalue sensitivity analysis (participation factors), the method reduces complex closed-loop pole placement to computationally efficient diagonal node ranking, offering an interpretable criterion for screening candidate intervention targets in dynamical networks.

Critical weaknesses
  1. There is a fundamental disconnect between the core identifiability theory and the real biological data experiments. The uniqueness guarantee in Proposition 1 relies entirely on the block-diagonal sparsity structure specific to coupled mRNA-protein networks (which eliminates 3n2−3n3n^2 - 3n degrees of freedom to make the Lyapunov system overdetermined), yet the sole biological validation in Section 6 (Paul et al., 2015) contains only mRNA measurements; without protein data, the theoretical block identifiability collapses, forcing the authors to manually pre-select 26 known hematopoietic transcription factors and impose an external Spearman-correlation threshold mask (while similarly feeding the ground-truth Jacobian support mask into the synthetic experiments in Appendix J.3).

  2. Modeling normal lineage differentiation clusters as quasi-stationary Gaussian fluctuations around local equilibria is biologically implausible, and the actual implementation deviates from the theoretical formulation. Rather than capturing a disease bifurcation process, the real-data experiment mean-centers and pools 15 distinct normal cell types across divergent branches (erythroid, myeloid, and lymphoid), erasing macroscopic developmental drift; moreover, whereas Theorem 1 relies on a temporal kernel KT((t−tk)/h)K_T((t - t_k)/h) to borrow strength across adjacent stages, the actual weight implementation in Appendix K.3 Equation (32) completely omits KTK_T, resulting in indiscriminate pooling that dilutes pre-bifurcation critical slowing-down signals.

  3. Non-parametric kernel conditional density estimation suffers from a severe curse of dimensionality in high-dimensional spaces, and the control formulation offers limited algorithmic novelty over the authors' prior work. Theorems 1 and 2 require h→0h \to 0 and NHh2n→∞N H h^{2n} \to \infty under a known isotropic Gaussian white noise assumption D=σ2I2nD = \sigma^2 I_{2n}, which not only contradicts single-cell transcriptional bursting and state-dependent count noise characteristics (Larsson et al., 2019; Sarkar & Stephens, 2021; Gorin et al., 2022), but also becomes statistically intractable due to exponential h2nh^{2n} decay even in the reduced 90-dimensional synthetic space and 50-dimensional real gene panel with only 375 total cells (forcing the authors to hard-code a global bandwidth h=1.0h = 1.0 on [−1,1][-1, 1]-scaled data in Appendix K.3); furthermore, the optimal and diagonal re-stabilization objectives and participation-factor criterion (Equation 10 and Lemma 1) are directly adopted from Yasukata et al. (2023) and Shen et al. (2024a, 2024b).

  4. The empirical evaluation is restricted to small-scale static node ranking, lacking closed-loop treatment verification and adequate baseline comparisons. Although motivated by early intervention in high-dimensional (n>104n > 10^4) single-cell systems, the experiments only evaluate Top-kk node identification accuracy in tens of dimensions without demonstrating that applying the synthesized feedback BθKθB_\theta K_\theta actually re-stabilizes trajectories or prevents bifurcation in the nonlinear system; additionally, the real-data experiment on the pre-filtered 50-gene panel includes no quantitative comparisons against baseline methods (such as scInTime, Shen et al., 2024b, or established single-cell perturbation tools like CellOracle and dynamo).

References

Questions for authors

see weaknesses

4 comments
AuthorsOpenReview ↗

Part 1/3: Closed-loop verification and quantitative comparisons

Part 1/3: Closed-loop verification and quantitative comparisons

Thank you for the comments and for emphasizing the importance of testing the selected interventions. We now apply the estimated targets to the true Jacobians, using the same fixed gain for every method.

Three estimated protein targets lower the synchronous spectral abscissa in every trial. Under the stated support prior and gain gi=−0.5g_i=-0.5, the abscissa falls from −0.2977-0.2977 to −0.5190-0.5190 in 100/100 trials at both N=50N=50 and N=70N=70 (Table 1):

Synchronous success (%)N=50N=50: l=1/3/6l=1/3/6N=70N=70: l=1/3/6l=1/3/6
Proposed0 / 100 / 1000 / 100 / 100
Single-snapshot covariance0 / 10 / 130 / 33 / 36
scInTime0 / 0 / 110 / 2 / 18
True κ\kappa0 / 100 / 1000 / 100 / 100

One target cannot lower the abscissa when two other modules retain the same critical eigenvalue. The zero entries at l=1l=1, including for true participation factors, agree with the revised coverage condition in Lemma 1. Full module coverage and successful control coincide in the reported trials.

The localized target is recovered and actuated without the oracle mask. Fitting the full biological block subspace, with 2,1602{,}160 free coefficients and no gene-specific regulatory mask, gives top-1 recovery of 88% and 94% at N=50N=50 and 7070, versus 12% and 60% for the single-snapshot fit and 0% for scInTime (Table 8). Actuating one estimated protein lowers the true abscissa in 88% and 94% of trials, versus 11% and 60% for the single-snapshot fit (Table 7). The all-node ranking and protein-restricted intervention endpoints are reported separately.

With the support prior, the proposed single target succeeds in every localized trial at both sample sizes, while degree succeeds in 39% and 34%. All three master proteins have out-degree 15, but only the critical hub changes the abscissa, from −0.1106-0.1106 to −0.3381-0.3381 (Tables 6-7). App. J.6 also reports mRNA-only interventions and a gain sweep. At gain magnitude 0.1, the true-projector first-order prediction differs from the achieved margin by less than 4% in both networks (Table 9).

Quantitative real-data comparison. We have added a new C7-to-C2 erythroid analysis using separate cells for prior training, validation and evaluation. In the post-hoc nuclear-transcript panel, an established reference regulator is ranked first in 29/100 evaluations, versus 15/100 for matched or likelihood-selected variance and 19/100 for reference-tuned variance. Prior out-degree and GENIE3 outgoing importance each give 0/100 for this endpoint. Mean non-negligible reference hits among the top five are 1.01, compared with 0.78, 0.80 and 0.81 for the variance controls (Table O.1). Part 2 specifies the design and interpretation; Table O.2 reports the panel comparisons.

Intervention scope and other tools. The closed-loop tests directly measure the spectral-abscissa criterion in Problem 1. Testing whether the same interventions prevent tipping along nonlinear Hill-SDE trajectories is a further evaluation. We identify this as future work, with the intervention time, parameter path and tipping criterion fixed in advance.

The added biological benchmarks compare variance and training-derived prior rankings. CellOracle infers context-specific regulatory networks and then predicts the expression and cell-identity response to a specified TF perturbation. Our method addresses an inverse problem: recovering continuous-time local dynamics from snapshot covariances, then using the estimated critical modes to select intervention targets. We evaluate those targets by their effect on the spectral abscissa under diagonal feedback. CellOracle's forward perturbation predictions do not directly provide this Jacobian-based identification and control test. The dynamo vector-field workflow uses RNA-velocity information, which is unavailable in this expression-only analysis.

Scale. The block fit has p=n2+3np=n^2+3n unknowns, or 2,1602{,}160 at n=45n=45. App. N reports approximately one hour on the specified laptop CPU for the block benchmark with 100 trials over seven sample sizes. This is a panel-scale implementation; genome-wide fitting would require additional sparsity and sparse/operator-based solvers.

AuthorsOpenReview ↗

Part 2/3: Identifiability, biological interpretation and the temporal estimator

Part 2/3: Identifiability, biological interpretation and the temporal estimator

The generic block experiment tests the structure used by Proposition 1. It removes the true regulatory support while keeping the mRNA-protein block constraints. The localized results above therefore establish target recovery without the oracle mask. At the population covariances of the true Jacobians, both synthetic restricted operators have full column rank, 2,1602{,}160. Table 8 also reports the synchronous generic-block results: its simultaneous critical modes are not resolved at the fixed validation-selected bandwidth. We distinguish this setting from the synchronous support-prior coverage experiment. Complete recovery of the critical gene's mRNA and protein states is reported separately: top-2 recovery is 20% and 40% at N=50N=50 and 7070, and 63% at N=100N=100.

The real-data section now evaluates regulator concordance. The mRNA-only reduction does not inherit Proposition 1's block identifiability. Under the stationary model with isotropic diffusion, the original Spearman mask has a 15-gene component with 137 unknowns but only 120 symmetric equations, giving total operator rank at most 172 for 196 unknowns (App. K.3). Positive empirical rank can result from cross-component covariances due to sampling variation or model mismatch. It does not establish identifiability under that model. The submitted inverse-covariance fit is retained as an explicitly labeled auxiliary result, with its prior, initialization and gene roles stated in Apps. K.3-K.4.

The new study follows the published C7-to-C2 erythroid lineage. Five stage-stratified splits separate 608 prior-training cells, 304 validation cells and 307 evaluation cells per split. GENIE3 uses only the prior-training cells, with all panel genes eligible as regulators. Model sparsity and bandwidth are selected by validation covariance likelihood. Variance uses matched bandwidths and independent validation tuning, including a reference-recovery selector. Each split contributes 20 evaluations with 25 cells per stage; the 100 evaluations are nested within the five splits.

Following the initial mitochondrial-dominance diagnostic, the nuclear-transcript sensitivity analysis removes only Atpase6, leaving 49 genes and the induced subnetwork of the training-derived prior. The fixed references are Gata1, Klf1, Gfi1b, Gata2 and Zfpm1. All 100 fitted systems are stable and their empirical operators have full column rank. Four reference genes appear among the top five with non-negligible scores: Zfpm1 in 38 evaluations, Gfi1b in 27, and Gata1/Klf1 in 18 each. Table O.1 gives the nuclear first-ranked-reference and top-five comparisons. Table O.2 includes the complete panel results; the corresponding full 50-gene dispersion panel gives 10/100 first-ranked references, versus 8/100 for reference-tuned variance.

The Paul data follow normal erythroid differentiation, so this biological comparison tests whether the method prioritizes established regulators of that process. The nuclear panel was examined as a post-hoc follow-up. Its reference genes come from the same literature used to construct the panel and provide an incomplete benchmark. Tables O.2 and O.3 report the complete panel comparisons. The synthetic experiments separately measure how the estimated interventions change local stability.

Temporal kernel and centering. All revised covariance-based synthetic fits use C^(h)=∑k=1Hwk(h)Sk,wk(h)∝KT((tH−tk)/h),\widehat C(h)=\sum_{k=1}^{H}w_k(h)S_k,\qquad w_k(h)\propto K_T((t_H-t_k)/h), where SkS_k is the unbiased covariance around the mean of the selected cells of episode kk (Eq. (7)). The kernel acts on the episode coordinate. Differences between episode means therefore do not enter the covariance, and Gaussian centering costs one degree of freedom without covariance bias (Lemma 2).

The submitted real-data fit omitted this temporal kernel and is not covered by Theorems 1-3. Its direct-estimator diagnostic uses the uniform-weight limit because the original mixed cell types have no single temporal ordering. The new lineage study supplies an explicit stage coordinate and uses the direct estimator. Table 4 distinguishes these procedures. The analysis assumes quasi-stationarity and approximately Gaussian local fluctuations.

AuthorsOpenReview ↗

Part 3/3: Statistical rate, diffusion assumptions and contribution

Part 3/3: Statistical rate, diffusion assumptions and contribution

The guarantee is derived from the original cells without state-space density estimation. The revised estimator smooths only over the episode coordinate, so the condition NHh2n→∞NHh^{2n}\to\infty no longer arises. Set dx=2nd_x=2n, u=log⁡(2dx(dx+1)/δ)u=\log(2d_x(d_x+1)/\delta), cC=max⁡k∥Ck∥2c_C=\max_k\lVert C_k\rVert_{2}, wmax=max⁡kwkw_{\mathrm{max}}=\max_k w_k and neff=(∑kwk2)−1n_{\mathrm{eff}}=(\sum_k w_k^2)^{-1}. Theorem 2 gives ϵ=dxcC[22u(N−1)neff+2uwmaxN−1]+LC∑kwk∣tH−tk∣.\epsilon=d_x c_C\left[2\sqrt{2}\sqrt{\frac{u}{(N-1)n_{\mathrm{eff}}}}+\frac{2u w_{\mathrm{max}}}{N-1}\right]+L_C\sum_k w_k|t_H-t_k|. With probability at least 1−δ1-\delta, ∥A^−A∥F≤8∥A∥Fϵσ⋆when ϵ≤σ⋆/4.\lVert\widehat{A}-A\rVert_{F}\le\frac{8\lVert A\rVert_{F}\epsilon}{\sigma_{\star}}\quad\text{when }\epsilon\le\sigma_{\star}/4.

The assumptions are explicit: stationary Gaussian cells, independent samples within and across episodes, temporal covariance regularity and a positive population restricted margin. Cross-episode independence is not inferred from destructive measurement. The bound separates sampling error and temporal bias and displays its dimension dependence. Its population constants are not replaced by empirical singular values to claim a data-dependent certificate.

At H=10H=10, the selected bandwidth h⋆=0.8h_\star=0.8 gives neff=7.29n_{\mathrm{eff}}=7.29 (Table 3). This bandwidth was fixed using an independent validation network before synthetic testing (App. F). For equispaced episodes, h→0h\to0 and Hh→∞Hh\to\infty give consistency for any fixed N≥2N\ge2. With dimension and population constants fixed, the leading stochastic term decreases with [(N−1)Hh]−1/2[(N-1)Hh]^{-1/2}, while temporal bias is O(LCh)O(L_Ch). Thus temporal pooling avoids the spatial bandwidth exponent while retaining an explicit bias-variance trade-off.

Known diffusion and biological noise. The regression accommodates any known constant diffusion covariance DD: its right-hand side is −vech(D)-\mathrm{vech}(D) and the restricted operator retains the same form. The covariance and matrix-estimation bounds continue under their stated assumptions and σ⋆>0\sigma_{\star}>0. Proposition 1 establishes generic identifiability for isotropic diffusion; for another known DD, the corresponding rank condition must be established separately. This extension concerns process diffusion. State-dependent bursting, unknown diffusion and count measurement noise require additional modeling and are outside the present guarantees.

Contribution relative to the cited control literature. We credit Yasukata et al. (2023) and Shen et al. (2024a, b) for the feedback objectives and simple-eigenvalue sensitivity criterion. Our contribution is the structured recovery of intervention-relevant spectral quantities from snapshot covariances, with identifiability and original-sample estimation bounds and their propagation to participation scores.

For repeated, semisimple critical eigenvalues, the revised Lemma 1 states the required coverage condition through Ω=Wd⊤BKVd,tr⁡Ω=∑igiκ(si).\Omega=W_d^\top BKV_d,\qquad\operatorname{tr}\Omega=\sum_i g_i\kappa(s_i). The eigenvalues of Ω\Omega determine the first-order shifts; its trace alone does not determine the largest real part. A strict first-order decrease requires max⁡Re⁡spec⁡(Ω)<0\max\operatorname{Re}\operatorname{spec}(\Omega)<0, which entails rank⁡(Wd⊤B)=q\operatorname{rank}(W_d^\top B)=q. An uncovered critical mode remains in the closed-loop spectrum for any gain. The synchronous interventions verify this coverage requirement, and the localized generic-block experiment verifies target selection without the oracle mask.

These results connect the stated structural and sampling assumptions to an implemented estimator and to measured closed-loop improvement. Sec. 6 and App. O provide a separate quantitative, exploratory assessment on biological snapshots.

AuthorsOpenReview ↗

Additional results for weaknesses 1 to 4

Additional results for weaknesses 1 to 4

We posted further experiments in the two comments "Additional experiments for all reviewers" at the top of the forum. The results that bear on your points, with the main numbers:

  • Weakness 1, identifiability without protein data (new test). The first step of the proof of Proposition 1 holds for any support that contains the diagonal, so it covers mRNA-only supports: full column rank at one stable matrix gives identifiability for almost every stable system on that support. Each of the five supports of the erythroid analysis has full column rank at 20 of 20 random stable matrices. The submitted Spearman mask does not, as reported in our response.
  • Weakness 2, pooling and the pre-bifurcation signal (new measurement). We measured how much pooling with the time kernel dilutes the critical signal. In the true covariances, the critical direction carries about 30 times the variance of the next direction at the pre-disease stage and about 5 times at an earlier stage. Pooling lowers these to about 7 and 2. Recovery follows the same pattern. At the pre-disease stage the method ranks a critical master first in 87% / 98% of localized trials and 44% / 73% of synchronous trials (N = 50 / 70), against 18% / 61% and 16% / 35% for the single snapshot. With 150 cells the synchronous value reaches 97%, against 48%. At earlier stages, where the pooled ratio is about 2, the single snapshot is the better choice, and the revised PDF states (Appendix P.1) that the pooled estimator is meant for the pre-disease stage. This extends our earlier explanation: simultaneous critical modes raise the number of cells needed, and the distance from the transition is the larger factor.
  • Weakness 3, dimension and sample size. As stated in our response, the estimator pools over time only, so no density is estimated in the 90-dimensional state space. The new experiments use it with 50 and 70 cells per episode in 90 dimensions, with 2,160 free coefficients.
  • Weakness 4, closed loop in the nonlinear system (new test). We applied the feedback to the nonlinear model with noise and moved the threshold 0.6 past the saddle-node. Without feedback, 0% of cells stay on the healthy branch. Feedback with gain -0.5 on the three critical master proteins keeps 100% of cells healthy in the synchronous network, and feedback on the two critical nodes does so in the localized network (on the master protein alone it takes gain -1.0). Feedback on the highest-variance nodes, on scInTime's most frequent pick, or on a master of equal degree in a dormant module leaves the system tipping (0%). The targets are the critical nodes, which the method identifies with the support prior. With targets estimated without it, the test succeeds whenever the selected proteins cover all critical modules, at the coverage rates given below.
  • Weakness 4, baselines (new baseline and benchmark). We added a hub baseline with no fitted model: rank the proteins by their total GENIE3 importance. In a new benchmark where the strongest hub is not the critical master, it stays at 0%, and our fit rises from 14% at N = 50 to 62% at N = 300 with no support information (36% to 77% with a learned support). In the original benchmarks the critical master is also the strongest hub, so the hub baseline finds it there in at least 94% of trials. Closed loop without the oracle support, at the pre-disease stage: one estimated protein lowers the dominant eigenvalue of the true localized system in 87% / 98% of trials (N = 50 / 70). The synchronous test is stricter: the three top-ranked proteins have to cover all three critical modules. With the block subspace they do so in 92% of trials at N = 200 and 95% at N = 300, against 24% and 56% for the single snapshot. With fewer cells the rates are 5%, 23%, 46% and 71% at N = 50, 70, 100 and 150 (single snapshot: 0%, 0%, 0% and 7%). With a learned support they do so in 65% at N = 50.

We hope these results address your concerns, and we would be grateful if you would consider raising your score.

Reviewer_33Z1OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes identifying a linearized mRNA-protein regulatory system from stationary snapshot covariances with known isotropic diffusion. It restricts the inverse Lyapunov problem to a biologically structured matrix subspace, gives a generic identifiability argument and perturbation bounds, and pools episodes through a conditional density estimator. Estimated spectral participation factors then rank candidate nodes for diagonal re-stabilization. Evaluation uses two 90-dimensional modular synthetic networks and a 50-gene analysis of the Paul et al. mouse hematopoiesis dataset. The connection between structured identification and intervention is valuable, but the stated control guarantee and the statistical-to-experimental connection have substantial unresolved problems.

Critical strengths
  1. The paper recognizes the identifiability problem and supplies a substantive structural argument. Equations 7–9 correctly make injectivity of the restricted Lyapunov map the relevant condition, rather than merely counting equations. Appendix E provides an explicit witness and an analytic zero-set argument within a specified stable connected component. The dependence of the subsequent least-squares perturbation bound on the smallest singular value is also informative.
  2. The target is motivated by control rather than covariance alone. For a simple isolated eigenvalue, the left/right eigenvector product is the relevant first-order sensitivity to a diagonal perturbation. The synthetic study includes both synchronous and localized criticality, 100 trials per sample size and uncertainty intervals (Section 5; Appendix J.5). This is a useful starting point for investigating how estimation affects intervention-target selection.
Critical weaknesses
  1. The extension from a simple eigenvalue to a repeated critical eigenspace does not establish re-stabilization. Page 7 states that replacing the eigenvector product by the spectral-projector diagonal leaves the subsequent statements unchanged. For a semisimple repeated eigenvalue, first-order eigenvalue shifts are the eigenvalues of the perturbation restricted to the critical subspace; the projector diagonal instead describes a contribution to their sum. In the paper’s three independent, simultaneously critical modules, acting on a master in only one module leaves the other two dominant eigenvalues unchanged. Thus including a key gene need not strictly reduce the spectral abscissa in Problem 1. Appendix I bounds projector estimation, which does not repair this control issue. Ranking accuracy is therefore not a demonstration of successful treatment in the principal synthetic settings.
  2. The finite-sample guarantee is not connected rigorously to the data estimator being evaluated. Appendix H assumes the fitted-density second moment is already close to the population covariance in Eq. (27), then bounds Monte Carlo error from draws from that fitted density. Eq. (27) includes estimation error from the original snapshots, not just smoothing bias; its finite-sample dependence on N, H, bandwidth and cross-episode dependence is not established. Assumption 1 explicitly imposes no cross-episode independence, and nonzero kernel weights alone do not imply NH effective independent observations. On the real data, Eq. (32) uses state-density weights over pooled centered observations, without the time kernel or fitted-density resampling analyzed in Sections 4 and F–H. Consequently, the stated guarantees cannot simply be transferred to the reported experiment.
Questions for authors
  1. Correctness: For the synchronous three-module example, how does Lemma 1 ensure a strict decrease in the largest real part when an intervention reaches only one critical module? Please state the required coverage/gain conditions and report the actual closed-loop spectral abscissa under the proposed selections. A projector-error bound alone does not answer this question.
  2. Statistical clarification: Can you bound Eq. (27) from the original snapshot sample sizes under explicit temporal sampling/dependence assumptions, and state which estimator is implemented in each experiment? In particular, how is Eq. (32) covered by the time-conditional estimator’s guarantees, and what numerical identifiability margin was obtained for the real-data fit?
  3. Validation clarification: Appendix J.3 says the ground-truth last-episode Jacobian supplies the reference support pattern. Which entries are revealed to each method, and how much target-ranking accuracy follows from this support alone? Please distinguish this oracle prior from the generic block constraints. For the real-data analysis, what evidence beyond known lineage markers supports a causal re-stabilization interpretation?
5 comments
AuthorsOpenReview ↗

Q1: Critical-mode coverage and the actual closed-loop response

Q1: Critical-mode coverage and the actual closed-loop response

Thank you for this question. We have added an explicit coverage analysis and evaluated the estimated interventions on the true Jacobians. With three estimated protein targets and the same fixed gain for every method, the proposed selections lower the true spectral abscissa from −0.2977-0.2977 to −0.5190-0.5190 in 100/100 trials at both N=50N=50 and N=70N=70. The revised Lemma 1 explains this result: all three simultaneously critical modules must be reached. Actuating one module leaves the other critical modes unaffected.

Coverage and gain conditions (Sec. 4; App. I.1). Write BK=∑i=1lgiesiesi⊤BK=\sum_{i=1}^l g_i e_{s_i}e_{s_i}^{\top} and scale the gains by τ>0\tau\gt0. Assume a positive real-part gap between the critical eigenvalue and the remaining spectrum.

  • For a simple eigenvalue, the perturbed eigenvalue is λd+τ∑igiκ(si)+O(τ2)\lambda_d+\tau\sum_i g_i\kappa(s_i)+O(\tau^2).
  • For a semisimple eigenvalue of multiplicity qq, take Wd⊤Vd=IqW_d^\top V_d=I_q and define Ω=Wd⊤BKVd\Omega=W_d^\top BKV_d. The nearby eigenvalues are λd+τωj+o(τ)\lambda_d+\tau\omega_j+o(\tau), where ωj∈spec(Ω)\omega_j\in\mathrm{spec}(\Omega). If max⁡jRe ωj<0\max_j\mathrm{Re}\,\omega_j\lt0, the spectral abscissa strictly decreases for sufficiently small τ\tau. Since rank Ω≤rank(Wd⊤B)≤l\mathrm{rank}\,\Omega\le\mathrm{rank}(W_d^\top B)\le l, this condition requires coverage, rank(Wd⊤B)=q\mathrm{rank}(W_d^\top B)=q, and hence l≥ql\ge q. The projector diagonal determines tr Ω=∑igiκ(si)\mathrm{tr}\,\Omega=\sum_i g_i\kappa(s_i); control depends on the largest real part of the reduced eigenvalues.
  • If a dominant left or right eigenvector vanishes on every actuated node, λd\lambda_d remains in the closed-loop spectrum for any gains, preventing a strict decrease.

In the synchronous network, q=R=3q=R=3 and Ω\Omega is diagonal in a basis supported on the individual modules. Each module must contain an actuated master with an inhibitory gain. The qq-actuator requirement applies to an exactly repeated eigenvalue. For diagonalizable AA, a conservative sufficient separation condition is that τ cond(V)max⁡i∣gi∣\tau\,\mathrm{cond}(V)\max_i|g_i| is less than half the real-part gap. The finite gains used below are checked by direct closed-loop evaluation.

Closed-loop results. Each method's top-ll proteins receive gi=−0.5g_i=-0.5 on the true last-episode Jacobian, whose real-part gap is 0.5160.516. All methods use the same 100 cell subsamples as Fig. 3. Success means an abscissa reduction exceeding 10−610^{-6} (Table 1).

SourceN=50N=50: l=1/3/6l=1/3/6 (%)N=70N=70: l=1/3/6l=1/3/6 (%)
Proposed0 / 100 / 1000 / 100 / 100
Single-snapshot (SOTA)0 / 10 / 130 / 33 / 36
scInTime0 / 0 / 110 / 2 / 18
Out-degree0 / 100 / 1000 / 100 / 100
Random0 / 0 / 00 / 0 / 1
True κ\kappa0 / 100 / 1000 / 100 / 100

Success coincides with full critical-module coverage in every trial. Direct actuation gives abscissae of −0.2977-0.2977 for P15P_{15}, −0.2977-0.2977 for P15+P30P_{15}+P_{30}, and −0.5190-0.5190 for P15+P30+P45P_{15}+P_{30}+P_{45} (Table 6). The proposed mean at N=50N=50, l=3l=3 is −0.519-0.519. Its mRNA-only selections also succeed in every trial for l=3,6l=3,6 at both sample sizes. At gain magnitude 0.10.1, predictions using the true projector are within 4% of achieved margins. Estimated-projector predictions are about 1.5 times the true-projector predictions (Table 9), so we use estimated projectors to select targets and report achieved margins separately.

Supplied information and degree control. This synchronous experiment receives q=3q=3 and the last-episode support (App. J.3). Out-degree also identifies its three master proteins; it is therefore a coverage/control test under that prior. The localized experiment separates sensitivity from degree: P15,P30,P45P_{15},P_{30},P_{45} all have out-degree 15, but only P15P_{15} lowers the abscissa, from −0.1106-0.1106 to −0.3381-0.3381. At N=50N=50, a single proposed target succeeds in 100/100 trials, versus 39/100 for degree with random tie-breaking (Table 7). Q3 gives the experiment without the oracle mask.

Manuscript changes: Sec. 4 and App. I.1 state and prove the multiplicity-aware lemma and gain/coverage conditions. Sec. 5 and App. J.6 report the actual closed-loop abscissae, both intervention modalities and the gain sweep (Tables 1, 6, 7, 9). App. B gives the covariance-subspace formulation for simultaneously critical modes. Fig. 3 assesses k=1,3,6k=1,3,6 recovery. App. I defines the matched spectral projector and its conditioning constant; Theorem 3 gives cθ=32c_\theta=32 for normal AA with q=1q=1.

AuthorsOpenReview ↗

Q2: Original-sample guarantee and the estimator used in each experiment

Q2: Original-sample guarantee and the estimator used in each experiment

We now give a bound for the implemented estimator in terms of the original snapshot sample sizes. All revised covariance-based synthetic fits use

C^(h)=∑k=1Hwk(h)Sk,\widehat C(h)=\sum_{k=1}^H w_k(h)S_k,

where SkS_k is the unbiased covariance of the selected NN cells, centered at their own episode mean, and wkw_k are normalized Gaussian time-kernel weights (Eq. (7)). The matrix estimate uses the same unweighted vech\mathrm{vech} least-squares fit as the analysis. No draws from a fitted density are required.

Original-sample bound (Sec. 4; Apps. C, H). We assume i.i.d. stationary Gaussian cells within an episode and independent samples across episodes. Cross-episode independence is an explicit sampling assumption; destructive measurement alone does not imply it. Suppose ∥Ck−CH∥F≤LC∣tk−tH∣\Vert C_k-C_H\Vert_{F}\le L_C|t_k-t_H|. Set neff=(∑kwk2)−1,wmax=max⁡kwk,cC=max⁡k∥Ck∥2,n_{\mathrm{eff}}=(\sum_k w_k^2)^{-1},\quad w_{\mathrm{max}}=\max_k w_k,\quad c_C=\max_k\Vert C_k\Vert_{2}, dx=2nd_x=2n and u=log⁡(2dx(dx+1)/δ)u=\log(2d_x(d_x+1)/\delta). With probability at least 1−δ1-\delta, ∥C^−CH∥F≤dxcC[22u(N−1)neff+2uwmaxN−1]+LC∑kwk∣tH−tk∣=:ϵ.\Vert\widehat{C}-C_H\Vert_{F}\le d_x c_C\left[2\sqrt{2}\sqrt{\frac{u}{(N-1)n_{\mathrm{eff}}}}+\frac{2u w_{\mathrm{max}}}{N-1}\right]+L_C\sum_k w_k|t_H-t_k|=:\epsilon. For the population margin σ⋆=σmin⁡(MS(CH))>0\sigma_{\star}=\sigma_{\min}(M_{\mathcal{S}}(C_H))>0, the same event gives ∥A^−A∥F≤8∥A∥Fϵ/σ⋆when ϵ≤σ⋆/4.\Vert\widehat{A}-A\Vert_{F}\le 8\Vert A\Vert_{F}\epsilon/\sigma_{\star}\quad\text{when }\epsilon\le\sigma_{\star}/4. The two terms separate sampling error and temporal drift. Lemma 2 shows that a mean-centered Gaussian episode covariance has the distribution of an average of N−1N-1 independent Gaussian outer products. Estimating the mean thus costs one degree of freedom without covariance bias. The effective sample count is (N−1)neff(N-1)n_{\mathrm{eff}}: at H=10H=10, neff=3.98n_{\mathrm{eff}}=3.98 for h=0.4h=0.4 and 7.297.29 for h=0.8h=0.8 (Table 3).

For fixed state dimension, a fixed covariance path, equispaced episodes and a Gaussian kernel, consistency follows with h→0h\to0, Hh→∞Hh\to\infty and any fixed N≥2N\ge2. For fixed HH, it follows with N→∞N\to\infty and hN→0h_N\to0. The direct estimator avoids state-space density estimation and its NHh2nNHh^{2n} condition.

Implementation. Appendix F specifies the independent validation network, bandwidth grid and fixed choice h⋆=0.8h_{\star}=0.8 for synthetic testing. The single-snapshot comparator is SHS_H, the h→0h\to0 limit. Table 4 lists the covariance estimator, solver and structural information used in each experiment; synthetic margins use the vech\mathrm{vech} operator. In Q1's synchronous support-prior experiment, three estimated protein targets lower the actual closed-loop abscissa in 100/100 trials at both N=50N=50 and N=70N=70. Q3 reports the localized experiment without the oracle mask.

Real-data estimator and margin. The submitted Eq. (32) uses state-density weights without a time kernel. Its matrix fit minimizes an auxiliary inverse-covariance objective initialized at the prior. App. K.3 and Table 4 distinguish that fit from the direct time-kernel/vech\mathrm{vech} estimator. Under the stationary model with isotropic diffusion on the submitted Spearman support, the restricted operator has rank at most 172 for 196 unknowns, so its population margin is zero. The 15-gene component alone has 137 unknowns and 120 symmetric equations.

For the released seed-42 state-weighted covariance, the empirical diagnostic is σ^⋆/∥C^∥2=2.39×10−3\widehat{\sigma}_{\star}/\Vert\widehat{C}\Vert_{2}=2.39\times 10^{-3}. In the reported 100-subsample direct-estimator diagnostic, the margin is zero in 52 subsamples and ranges from 4.3×10−34.3\times 10^{-3} to 1.7×10−21.7\times 10^{-2} in the other 48. Cross-component empirical covariances can produce a positive diagnostic through sampling variation or model mismatch. Empirical singular values therefore describe conditioning; they do not certify population identifiability. The bound applies to the revised estimator under its stated assumptions and does not certify the submitted Eq. (32) fit. The new ordered-lineage study uses the direct estimator and is reported in App. O.

Manuscript changes: Sec. 4 gives the direct estimator and explicit assumptions. Apps. C, G, H provide the centering, consistency and concentration proofs; App. F and Table 4 specify the implementation and bandwidth selection. App. K.3 reports the original fit and its numerical/structural diagnostics. App. O gives the new biological comparison.

AuthorsOpenReview ↗

Q3 (Part 1/2): Structural information and support-only controls

Q3 (Part 1/2): Structural information and support-only controls

We thank the reviewer for asking us to distinguish the oracle support from the biological block constraints. The localized target is recovered without the oracle mask: fitting all 2,160 free block coefficients gives top-1 recovery of 88% at N=50N=50 and 94% at N=70N=70, compared with 12% and 60% for the single-snapshot comparator and 0% for scInTime (App. J.7, Table 8). Actuating one estimated protein target lowers the true Jacobian's spectral abscissa in 88% and 94% of trials (Table 7); two protein targets increase these rates to 98% and 100% (App. J.6). All-node ranking and protein-restricted intervention are separate endpoints. The principal result is recovery of an effective intervention target. Recovery of the complete two-state set (the critical gene's mRNA and protein) is 20% and 40%, respectively.

Information supplied. In the oracle experiments, both covariance methods receive the 180 nonzero positions of the last-episode Jacobian: 90 diagonal entries, 45 gene-matched translation entries and 45 regulatory entries. The true coefficient values are not supplied. The generic block fit uses only Eq. (4): the diagonals of Dr,Dp,GD_r,D_p,G and every entry of QQ are free. Thus it supplies no information about which protein regulates which gene. Both covariance methods receive the same structural information and known diffusion; scInTime infers its networks from snapshots. The multiplicity q=3q=3 is supplied in the synchronous tests.

What the support reveals. In the synchronous network, a support-only graph or spectral rule identifies the six feedback-cycle states, and out-degree identifies the three master proteins. This experiment tests eigenspace estimation, critical-module coverage and closed-loop improvement under that prior. The localized block experiment tests discovery without the mask. At the fixed validation-selected bandwidth, the synchronous block fit has 2% top-1 recovery at both N=50N=50 and N=70N=70 (Table 8); the discovery result is therefore specific to the localized setting. App. J.7 also verifies full column rank 2,160 of the population restricted operators at both simulated last-episode Jacobians, separately from their finite-sample performance.

Equal-degree control. In the localized support-prior experiment, P15,P30,P45P_{15},P_{30},P_{45} all have out-degree 15, but only P15P_{15} lowers the abscissa, from −0.1106-0.1106 to −0.3381-0.3381 at gain −0.5-0.5 (Table 6). Degree with random tie-breaking has expected top-1 recovery 1/31/3 and observed recovery 39/100 at N=50N=50 and 34/100 at N=70N=70, compared with 100/100 for the proposed method (Table 7). Equal degree therefore does not explain this dynamical ranking.

Original real-data prior. Ranking the submitted Spearman prior by out-degree gives four fixed genes from the submitted real-data table, with Gata1 and Sfpi1 tied fifth: expected overlap 4.5/5. A prior-only projector recovers 4/5. App. K.3 reports these prior controls alongside the auxiliary fitting objective and initialization. To test added value directly, the new biological comparison uses training-derived priors, prior-only rankings and matched variance controls (Part 2).

Manuscript changes: Sec. 5 distinguishes discovery under the generic block constraints from coverage under oracle support. App. J.3 states the supplied entries; App. J.7 and Table 8 report the block experiment and population-rank check. Apps. J.6 and K.3 give the synthetic equal-degree and original real-prior controls. App. E includes the connectedness argument establishing generic identifiability throughout the stable model class.

Reviewer_cSmxOpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
1(−1)
Read review
Paper summary

This paper addresses the problem of choosing intervention targets that prevent a gene regulatory network from undergoing a disease transition, using only single-cell snapshot data, in which each cell is measured once and the data come as unordered populations at successive stages rather than as trajectories. Building on Dynamical Network Biomarker theory, which detects the pre-disease stage from rising fluctuations, the authors linearize an mRNA–protein SDE around a quasi-stationary equilibrium and identify the system matrix from the stationary covariance through the Lyapunov equation. They restrict the unknowns to a structured subspace so that the problem becomes well posed. Since a single snapshot often has fewer cells than the state dimension, they pool cells across episodes with a kernel estimator. Proposition 1 shows that identifiability holds generically, with a computable certificate σ⋆\sigma_\star, and Theorems 1 to 3 give consistency and finite-sample bounds on the estimated matrix and on the dominant spectral projector, whose diagonal gives the participation factors κ\kappa used to rank genes. The method is evaluated on two 90-dimensional synthetic modular networks against a single-snapshot covariance method and scInTime, and on a 50-gene panel from the Paul et al. (2015) hematopoiesis dataset.

Critical strengths

The paper turns the snapshot identification problem into a well-posed one with a guarantee that can be checked on data. The structural zeros of the mRNA–protein model supply the constraints the symmetric Lyapunov equation lacks, the certificate σ⋆\sigma_\star can be computed from the estimated covariance and enters the error bounds of Theorems 2 and 3, and Proposition 1 reduces generic identifiability to a single explicit witness through analyticity, which makes the proof easy to verify. Theorem 3 is proved through the Riesz projector rather than a single eigenvector, so the analysis covers the synchronous case in which several modules become critical together. This is the main reason I support acceptance, since it connects snapshot data to intervention targets with an error guarantee at each step.

The synthetic experiments test the estimator the theory analyzes. The supplementary benchmark solves the linear least-squares problem of Section 4 and uses a strict criterion that every top-kk node must be a master, with 100 trials, Wilson intervals and paired McNemar tests, and its settings match Section 5 and Appendix J. The comparison between pooled and single-snapshot covariance is therefore controlled and supports the claim that pooling improves sample efficiency in the regime with fewer cells than dimensions.

Critical weaknesses

The real-data result is not produced by the estimator the paper analyzes. Section 6 (lines 453 to 459) states that the participation factors ranked in Table 1 come from solving the approximate identification problem, which Section 4 (lines 249 to 261) defines as a structured least-squares problem, but 06_estimate_regulatory_node.py minimizes the squared Frobenius distance between C^\hat{C} and C(A)C(A) with L-BFGS-B, starting the first of five restarts from the Spearman prior and the others from random values with its signs. The authors’ own docstring for reconstruct_jacobian_and_kappa_lstsq in estimate_regulatory_nodes.py states that the error bounds do not apply to this objective and that starting from the prior lets it influence the result. The run-time identifiability check that Appendix K.3 (lines 1546 to 1553) describes as a safeguard, reporting the certificate σ⋆\sigma_\star relative to the norm of C^\hat{C} for the real data, is not computed in the code or reported in the paper. In addition, the real-data README reports that the five genes with the highest out-degree in the Spearman prior are the same five genes as in Table 1. Since that prior sets both the support and the starting point, this agreement suggests the ranking largely reflects the prior’s hub structure, consistent with the presence of Mpo, Car2 and Elane, which Appendix K.4 describes as a granulocyte marker, a downstream erythroid effector and a neutrophil protease, among genes that Section 6 (lines 460 to 462) calls independently validated master regulators. Table 1 also comes from a single random subsample. The paper presents this experiment as biological validation, so this weakness limits what the paper shows beyond simulation.

The synthetic advantage over scInTime relies on an oracle support. Following Appendix J.3, both covariance methods receive the support of the last episode’s true Jacobian, which in the supplied data has 180 nonzero entries against n2+3n=2,160n^2 + 3n = 2{,}160 for the subspace of Eq. 7, while scInTime receives no structural information. The docstring of generate_snapshots_indep.py states that this support alone identifies the masters, since each master protein regulates 15 genes while every other protein regulates none, so the reported margin over scInTime (96% against 34% top-1 in Section 5) partly reflects this information. The benchmark code includes a GENIE3 option that replaces the oracle support with one inferred from data, but the genie3_prior module it imports is not in the supplement, and the paper does not report a result with it. The SOTA baseline is also implemented as the single-snapshot limit of the authors’ own estimator, with bandwidth 0.001 against 0.4 and the same solver and oracle support, which the paper should state, since Section 5 presents it only as the method of Shen et al. (2024b). As a result, the paper does not yet show an advantage over an external method under comparable information.

Questions for authors

Clarifications:

  1. What does Table 1 look like when computed with the least-squares solver of Section 4 on the support of the Spearman prior, and what is the identifiability margin σ⋆\sigma_\star relative to the norm of C^\hat{C} for that problem?
  2. How does the proposed method compare with scInTime in Section 5 when it uses a support inferred from data, such as the GENIE3 option in benchmark.py, whose genie3_prior module is missing from the supplement?
  3. How stable are the top five genes in Table 1 across random subsamples of 25 cells per episode?

Suggestions for future work:

Applying the designed feedback to the nonlinear Hill-function SDE near the bifurcation would show whether it prevents tipping. A dataset that follows an actual disease transition, rather than normal differentiation, would match the paper’s motivation more closely.

8 comments
AuthorsOpenReview ↗

Part 1/3: Comparable information and target discovery without the oracle mask

Part 1/3: Comparable information and target discovery without the oracle mask

Thank you for examining both the manuscript and the implementation. We have added a test of target discovery without the true regulatory support and evaluated the estimated targets on the true Jacobians. We also clarify what information each method receives.

The localized target is recovered without the oracle mask. The block-subspace experiment imposes only the biological constraints in Eq. (4): the diagonals of Dr,Dp,GD_r,D_p,G and every entry of QQ are free. The fit therefore has 2,1602{,}160 unknowns and receives no gene-specific regulatory edge information. Both covariance methods use the same cell subsamples, solver and block constraints (App. J.7):

Localized network, top-1 recovery (%)N=50N=50N=70N=70
Proposed, block subspace8894
Single-snapshot, same block subspace1260
scInTime00

Actuating one estimated protein target lowers the true Jacobian's spectral abscissa in 88% and 94% of trials, compared with 11% and 60% for the single-snapshot fit (Table 7). The two-protein check in App. J.6 gives 98% and 100%. All-node ranking and protein-restricted intervention are distinct endpoints. The principal result is selection of an effective intervention target. Complete top-two recovery of the critical gene's mRNA and protein states is 20% and 40%, increasing to 63% at N=100N=100 (Table 8). The population restricted operators have full column rank, 2,1602{,}160, at the true Jacobians of both synthetic networks.

Information supplied. In the support-prior experiments, both covariance methods receive the 180180 nonzero positions of the last-episode Jacobian: 9090 diagonal, 4545 gene-matched translation and 4545 regulatory entries. Their values are not supplied. scInTime infers its networks from the snapshots. Table 4 lists the estimator, solver and structural information for each experiment. The synchronous estimator is also supplied the multiplicity q=3q=3.

The synchronous support identifies the three master proteins by out-degree and the six feedback-cycle states by a support-only graph or spectral rule. This experiment therefore evaluates eigenspace estimation and critical-module coverage under the stated prior. Table 8 also reports the generic-block synchronous experiment: top-1 recovery is at most 17% at the fixed validation-selected bandwidth. The localized generic-block experiment supplies the successful discovery result without that mask.

Equal-degree control. In the localized network, P15,P30,P45P_{15},P_{30},P_{45} all have out-degree 1515, while only P15P_{15} lowers the abscissa, from −0.1106-0.1106 to −0.3381-0.3381 at gain −0.5-0.5. With the support prior, the proposed single target succeeds in 100% of trials; degree with random tie-breaking succeeds in 39% at N=50N=50 and 34% at N=70N=70 (Tables 6-7). Equal degree does not explain the dynamical ranking.

Single-snapshot and learned-support comparisons. Sec. 5 identifies SOTA as the single-snapshot limit h→0h\to0 of Eq. (7), using the same regression and structural information. The proposed bandwidth h⋆=0.8h_\star=0.8 was selected on an independent validation network and frozen before testing (App. F). Under the synchronous support prior at N=50N=50, top-1 recovery is 100% for the proposed method, 56% for the single-snapshot fit and 34% for scInTime.

Using the GENIE3-inferred support gives localized top-1 recovery of 7% and 9% at N=50N=50 and 7070, and 10% and 12% for the single-snapshot fit on the same support, versus 0% for scInTime (App. J.7). A learned mask can exclude true edges and thereby misspecify the regression; this is one possible source of the loss. The block experiment avoids imposing gene-specific regulatory zeros and provides the stronger non-oracle comparison.

Manuscript changes: Sec. 5 defines the information and single-snapshot comparator; Apps. J.6-J.7 report closed-loop, equal-degree, learned-support and generic-block results; App. F and Table 4 specify the implemented estimators and validation-selected bandwidth.

AuthorsOpenReview ↗

## Part 2/3: Real-data estimator, identifiability and subsampling stability

Part 2/3: Real-data estimator, identifiability and subsampling stability

The new erythroid analysis ranks a reference regulator first in 29/100 evaluations, versus 15/100 for matched or likelihood-selected variance and 19/100 for reference-tuned variance. Sec. 6 leads with this quantitative comparison; App. K documents the submitted auxiliary fit and its conditioning separately.

Submitted fit and structural identifiability. The submitted five-gene ranking used state-space kernel weights, h=1h=1 on [−1,1][-1,1]-scaled data, no temporal kernel, and the objective ∥C^−C(A)∥F2\|\widehat C-C(A)\|_F^2, minimized by L-BFGS-B with prior-based initialization. App. K.3 describes this procedure explicitly; Theorems 1-3 are not invoked for that auxiliary fit.

The Spearman support contains 196196 free entries in components of 1515, 66, 22 and 22 genes, plus 2525 isolated genes. Under the stationary model with isotropic diffusion, cross-component population covariances vanish. The 1515-gene component alone has 137137 unknowns but only 120120 symmetric Lyapunov equations, giving total operator rank at most 172172. Its population identifiability margin is therefore zero. Proposition 1 concerns the mRNA-protein block subspace and does not establish identifiability for this mask.

Empirical margins and rankings. In the reported 100-draw experiment with 2525 cells per type, 5252 draws contain a zero-variance gene and have zero empirical margin. In the remaining 4848, the relative margin σ^⋆/∥C^∥2\widehat\sigma_\star/\|\widehat C\|_2 ranges from 4.3×10−34.3\times10^{-3} to 1.7×10−21.7\times10^{-2}, with median 1.1×10−21.1\times10^{-2}. For the released seed-42 state-weighted covariance, the empirical vech diagnostic is 2.39×10−32.39\times10^{-3}. Cross-component computed covariances can arise from sampling variation or model mismatch; a positive empirical margin is a conditioning diagnostic, not a population certificate.

The additional direct least-squares checks also distinguish material participation from numerical ties. Non-negligible participation concentrates on Malat1; completing a five-gene list with zero or near-zero scores does not identify four additional targets. We therefore report material-score counts and do not interpret a repeated raw top-five list as stable recovery of five regulatory nodes.

Prior-only controls. Out-degree on the submitted Spearman prior selects Mpo, Car2, Elane and Klf1, with Gata1 and Sfpi1 tied fifth, giving expected overlap 4.5/54.5/5 with the submitted real-data ranking. A prior-only projector recovers 4/54/5. The auxiliary result is consequently reported with its prior and initialization, rather than as independent evidence for the Sec. 4 estimator. App. K.4 distinguishes upstream regulators Klf1/Gata1 from erythroid and myeloid effectors Car2/Mpo/Elane.

New biological comparison. Sec. 6 and App. O examine the ordered C7-to-C2 erythroid lineage. Five stage-stratified splits use 608 cells for prior training, 304 for validation and 307 for evaluation per split. All panel genes are eligible GENIE3 regulators. Sparsity and bandwidth are selected by validation covariance likelihood. Variance uses matched bandwidths and independent validation tuning, including a reference-recovery selector. Each split has 20 evaluations with 25 cells per stage; the 100 evaluations are nested within these five splits.

A follow-up nuclear-transcript analysis removes only Atpase6, leaving 49 genes and the induced subnetwork of the training-derived prior. The fixed reference set is Gata1, Klf1, Gfi1b, Gata2, Zfpm1 (Table O.1):

MethodReference regulator ranked first /100
Proposed29
Matched or likelihood-selected variance15
Reference-tuned variance19
Prior out-degree / GENIE3 outgoing importance0 / 0

Mean non-negligible reference hits among the top five are 1.01, compared with 0.78, 0.80 and 0.81 for the variance controls (Table O.1). Reference genes with non-negligible top-five scores include Zfpm1 in 38 evaluations, Gfi1b in 27, and Gata1/Klf1 in 18 each. All 100 fitted systems are stable and their empirical operators have full column rank.

After the initial mitochondrial-dominance check, we examined the nuclear panel as a post-hoc follow-up. Table O.2 retains the complete panel comparison, including the 50-gene dispersion panel, which gives 10/100 first-ranked references versus 8/100 for reference-tuned variance. These comparisons assess recovery of established regulators during normal erythroid differentiation.

Manuscript changes: Sec. 6 presents the new concordance study; Apps. K.3-K.4 retain and label the auxiliary fit, diagnostics and gene roles; App. O gives the split protocol, controls, material scores and panel comparisons.

AuthorsOpenReview ↗

Part 3/3: Statistical guarantee and intervention scope

Part 3/3: Statistical guarantee and intervention scope

The guarantee now applies directly to the original snapshots. All revised covariance-based synthetic fits use the direct temporal covariance C^(h)=∑kwk(h)Sk\widehat C(h)=\sum_k w_k(h)S_k, where SkS_k is the unbiased covariance of the selected cells centered at their own episode mean. Theorem 2 bounds this estimator using an effective count (N−1)neff(h)(N-1)n_{\mathrm{eff}}(h) and an explicit temporal-bias term. It assumes independent stationary Gaussian samples within and across episodes; cross-episode independence is stated as an assumption, not inferred from destructive measurement. Table 4 separates this estimator from the submitted auxiliary real-data fit.

The bound depends on population quantities. We therefore report empirical singular values as conditioning diagnostics and state the positive population-margin condition explicitly. For equispaced episodes and a Gaussian temporal kernel, h→0h\to0 and Hh→∞Hh\to\infty give consistency for fixed N≥2N\ge2. The estimator does not estimate a state-space density and requires no synthetic resampling from one.

Actual closed-loop verification. Under the synchronous support prior, three estimated protein targets with the common gain −0.5-0.5 lower the abscissa from −0.2977-0.2977 to −0.5190-0.5190 in 100/100 trials at both N=50N=50 and 7070 (Table 1). App. J.6 also tests the localized network, modality restrictions, coverage and gain dependence. These interventions evaluate the spectral-abscissa criterion in Problem 1 directly, rather than using ranking accuracy as its substitute.

Thank you also for suggesting a test of tipping prevention in the nonlinear Hill-function SDE. Our closed-loop experiments measure the change in local stability on the true Jacobians. A trajectory-level test would also require fixed intervention times, a specified parameter path and a criterion for tipping. We have added this further evaluation to the future-work discussion.

The revisions thus provide a direct original-sample guarantee, target discovery under generic biological constraints, and closed-loop verification under a stated intervention criterion. The real-data comparison supplies a separate, reproducible assessment of regulator prioritization.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Additional experiments for all reviewers (1/2): recovery depends on the distance from the tipping point

Additional experiments for all reviewers (1/2): recovery depends on the distance from the tipping point

We ran four further experiments after posting our responses. The main results:

  1. Distance from the tipping point (this comment). At the pre-disease stage, immediately before the transition, the proposed method with no support information ranks a critical master first in 87% / 98% of localized trials and 44% / 73% of synchronous trials at N = 50 / 70 (single snapshot: 18% / 61% and 16% / 35%). The synchronous value reaches 100% at N = 300. The 2% on the shipped synchronous benchmark is explained by its last episode being further from the transition.
  2. Hubs and critical nodes (part 2/2). In a new benchmark where the strongest hub is not the critical master, the hub baseline and scInTime stay at 0%, and our fit rises from 14% at N = 50 to 62% at N = 300 with no support information. In the original benchmarks the critical master is also the strongest hub, so the hub baseline finds it too.
  3. Closed loop in the nonlinear model (part 2/2). With the threshold moved past the saddle-node, feedback on the critical targets keeps 100% of cells on the healthy branch. Without feedback, 0% stay.
  4. Identifiability on real data (part 2/2). The supports of the erythroid analysis pass an identifiability test that follows from the proof of Proposition 1.

The numbers reported in our responses are unchanged. One explanation given there is extended below. Scripts and raw outputs are in the updated supplement.

Recovery without the oracle support depends on the distance from the tipping point. In the paper's setting the last episode is the pre-disease stage, immediately before the transition (Sec. 3). The two shipped benchmarks differ in how close their last episode is to it: its dominant eigenvalue is -0.11 in the localized network and -0.30 in the synchronous one. We generated both network types at four distances (two parameter draws each, 1,000 cells per episode) and applied the estimator of Table 8: block subspace, no support information, bandwidth 0.8, nothing tuned. Share of trials in which the top-ranked node is a critical master (%), at N = 50 / N = 70, with 50 trials on each of the two networks:

Last-episode eigenvalue-0.30-0.21-0.15-0.10
Localized, proposed5 / 320 / 3166 / 7887 / 98
Localized, single snapshot20 / 3323 / 428 / 4118 / 61
Localized, scInTime4 / 00 / 00 / 00 / 0
Synchronous, proposed1 / 114 / 2227 / 4144 / 73
Synchronous, single snapshot35 / 4216 / 3915 / 3816 / 35
Synchronous, scInTime24 / 2621 / 2418 / 2713 / 16

The first column is the distance of the shipped synchronous benchmark (Table 8: 2 / 2 for the proposed method, 34 / 43 for scInTime). The last column is that of the shipped localized benchmark (Table 8: 88 / 94). The code behind Table 8 and a separate re-implementation agree on every cell within 3 points. The two parameter draws agree: 86% and 88% (localized) and 40% and 48% (synchronous) in the last column at N = 50.

  • At the pre-disease stage (last column), the proposed method is above the single snapshot and scInTime in both network types.
  • The pooled estimator is meant for the pre-disease stage, and the revised PDF states this (Appendix P.1). At earlier stages (first column) the single snapshot is the better choice.
  • This extends the explanation in our responses. Simultaneous critical modes raise the number of cells needed (44% against 87% at N = 50, both 97% at N = 150). The distance from the transition is the larger factor: at the same distance the localized network gives the same value as the synchronous one.

The reason is signal strength. In the true last-episode covariance the critical direction carries about 5 times the variance of the next direction at -0.30 and about 30 times at -0.10, and pooling lowers these ratios to about 2 and 7. Applied to the true pooled covariances of all 16 networks, the fit ranks a critical master first every time, so the first-column values are a small-sample effect.

With more cells the last-column values rise further. In the synchronous network the proposed method reaches 85%, 97% and 100% at N = 100, 150 and 300, against 38%, 48% and 86% for the single snapshot. Its three top-ranked proteins then lower the dominant eigenvalue of the true system in 46%, 71% and 95% of trials (single snapshot: 0%, 7% and 56%). In the localized network both methods are at 93% or above from N = 100.

The stage can be read from the snapshots alone. The largest eigenvalue of the last-episode sample covariance, divided by that of the first episode, has median 12 to 15 in the last column and about 2 in the first (N = 50, no overlap in 100 trials). On the shipped benchmarks it is 10 (localized) and 2.4 (synchronous), again with no overlap.

AuthorsOpenReview ↗

Additional experiments for all reviewers (2/2): a hub baseline, a new benchmark, a nonlinear closed-loop test, and identifiability on real data

Additional experiments for all reviewers (2/2): a hub baseline, a new benchmark, a nonlinear closed-loop test, and identifiability on real data

A hub baseline. In the original benchmarks the critical master is also the strongest hub in the data, so a hub score with no model of the dynamics finds it too. We added this baseline: train GENIE3 on the selected cells, centred per episode, with the proteins as candidate regulators, and rank the proteins by their total importance. It ranks a critical master first in 94-100% of trials on the shipped localized network (N = 30 to 70), in 97-100% on the shipped synchronous network (N = 50 and 70), and in 84-100% on the networks of part 1/2 (N = 50). These forests differ from the GENIE3 run of App. J.7, which pooled the cells without centring. To test whether a method detects critical dynamics rather than hubs, we added a benchmark in which the two differ. The equal-degree control in our responses compared graph degree on the oracle support and is unchanged.

A benchmark that separates hubs from critical nodes. A second protein regulates 14 genes more than four times as strongly as the critical master but has no feedback loop, so it takes no part in the critical mode. The design was fixed before any run and nothing was tuned on it. Share of trials in which the top-ranked node is the critical master (%):

N = 50N = 100N = 300
GENIE3 importance, no fit000
Protein variance, no fit000
Out-degree in the learned mask, no fit333227
scInTime000
Single snapshot, learned support183564
Proposed, learned support364977
Single snapshot, block subspace81939
Proposed, block subspace143762

The no-fit rankings and scInTime do not improve with more cells. The fitted method does: it reaches 62% (block subspace) and 77% (learned support) at N = 300, and in the block subspace pooling is above the single snapshot at every sample size. Rows that use GENIE3 (out-degree and learned support) vary by up to about 10 points with the random seed of the forests. The block-subspace rows do not use it, and the two implementations agree on them within 5 points.

A first closed-loop test in the nonlinear model. The targets are the critical nodes, which the method identifies with the support prior (Tables 1 and 7). In the nonlinear model with noise we move the threshold 0.6 past the saddle-node. Without feedback, 0% of cells stay on the healthy branch. In the synchronous network, feedback with gain -0.5 on the three critical master proteins keeps 100% of cells healthy. Feedback on one of them keeps only its own module, as the revised Lemma 1 requires. In the localized network, feedback with gain -0.5 on the two critical nodes keeps 100% of cells healthy; on the master protein alone this takes gain -1.0. Feedback on the highest-variance nodes, on scInTime's most frequent pick, or on a master of equal degree in a dormant module leaves the system tipping (0%). Each condition uses 500 cells. With targets estimated without the support prior, the test succeeds whenever the selected proteins cover all critical modules; part 1/2 gives those coverage rates.

Identifiability of the supports used on real data. The first step of the proof of Proposition 1 holds for any support that contains the diagonal, because the path from any stable matrix A to -I, (1 - t) A - t I, stays on the support and stays stable. Full column rank of the identification operator at one stable matrix on a support therefore gives identifiability for almost every stable system on that support. This covers mRNA-only supports. We tested 20 random stable matrices per support. Each of the five erythroid supports has full column rank at all 20 (relative margin 0.02-0.05). The test also flags the submitted Spearman support, which has rank 172 of 196 at all 20, in line with the count in our responses. All 100 erythroid fits have full column rank. Their relative margin has median 0.065; 85 are above 0.01, and the other 15 come from one split (smallest 0.000002).

Changes to the paper. The revised PDF contains these experiments in Appendix P (Tables P.1-P.4), with the scope of the pooled estimator and the sample sizes of the new benchmark.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.