2608.09722-Decoding Gene Regulatory Networks from Single-Cell RNA Velocity

Introduction to the show: ident: Math Radio. The week's best mathematics papers, argued over on air.

Ada: Next we'll be talking about the paper "Decoding Gene Regulatory Networks from Single-Cell RNA Velocity".

Emmy: The paper was written by Lingqi Meng and Shiruo Wang from Yanbian University and State University of New York at Buffalo.

Ada: Stay tuned as we take you through the paper and discuss its implications.

Paper summary: ADA: A fresh arXiv paper landed this week that I couldn't put down, from Lingqi Meng at Yanbian University and Shiruo Wang at SUNY Buffalo, and today we're going to follow it page by page. It asks whether you can recover a gene regulatory network from single-cell RNA velocity data, given that every cell is measured once and then destroyed in the process.

EMMY: That destructive measurement is what makes the problem so strange. You get a snapshot of unspliced and spliced RNA, which gives you a local arrow for how each gene is changing, but the network behind those arrows is hidden. The paper treats that recovery as a sparse dynamical inverse problem and then proves when it's actually possible.

RAMAN: And the first big result is a negative one. Along a single unperturbed trajectory, distinct sparse regulatory networks can generate exactly the same observable RNA dynamics. The ambiguity is constructive, not a numerical artifact, so no estimator, however clever, can break it.

ADA: Then the paper flips the situation around. Controlled perturbations of the kind Perturb-seq produces enlarge the space of observable regulator trajectories, and identifiability returns exactly when a certain aggregated information matrix becomes positive definite.

EMMY: I love that there's a clean geometric core underneath. Identifiability comes down to whether the span of some augmented feature vectors fills the whole parameter space, and a one-dimensional trajectory can't do that on its own. You need the perturbations to stir up enough directions.

RAMAN: And they don't stop at theory. They build an integral sparse estimator that avoids numerical differentiation, prove finite-sample recovery guarantees that separate sampling noise from systematic errors, and validate the whole construction by predicting responses to perturbations the model never saw during fitting.

ADA: The practical message lands hard: unperturbed data has a mathematical ceiling, so perturbation design belongs inside the theory, not just in the lab protocol.

EMMY: And the theory quantifies which errors shrink as you collect more cells and which ones persist no matter how deeply you sequence.

RAMAN: I'm most curious to see the exact equations they start from, because the nonidentifiability proof has to grow out of the model structure itself.

ADA: Then let's turn to page one and see how they set it up.

Page 1 of the paper: ADA: The first page introduces RNA velocity the way anyone in single-cell biology knows it. La Manno and colleagues showed that the ratio of unspliced to spliced messenger RNA carries directional information about cell-state transitions, which turns static snapshots into local dynamics.

EMMY: But the paper immediately pushes past that. Velocity tells you how individual genes are changing, while the interesting biology sits one level higher — which transcription factors activate or repress which downstream genes. Recovering that regulatory network is a different kind of problem.

RAMAN: What struck me is how it positions the existing methods. Aubin-Frankowski and Vert used linear differential equations, Scribe coupled expression dynamics with velocity, and TFvelo brought regulatory structure into velocity estimation itself. These show velocity improves inference, but nobody had established the mathematical limits — whether the network is uniquely determined at all.

ADA: The page lays out three objectives that organize the whole paper. Characterize when unperturbed observations fail. Determine how perturbations restore identifiability. And construct a statistically stable estimator with guarantees that account for real-world uncertainty, including latent time and model error.

EMMY: I also appreciate the connection to classical systems theory. Structural identifiability, observability, persistent excitation — these are mature ideas from dynamical systems and control, and the paper treats single-cell data as a new proving ground for that old toolbox.

RAMAN: The phrase that stuck with me is "constructive rather than merely local." They don't just show a Jacobian is singular at one point. They build two distinct sparse networks that generate identical observable trajectories, which is a much stronger and more convincing failure.

ADA: And the introduction ends by promising that inferred networks will be judged by predicting cellular responses under held-out perturbations. That's a harder test than matching known edges, and it connects the inverse theory back to something measurable in the lab.

EMMY: So the promises are clear, and the model formulation has to explain where the ambiguity comes from. The next page should give us the equations.

Page 3 of the paper: ADA: This page builds the model, and it starts with two variables per gene. Unspliced RNA u and spliced RNA s. The dynamics couple them: transcription produces unspliced RNA at a rate that depends on the current regulator state, splicing converts unspliced into spliced, and degradation removes spliced RNA.

EMMY: And the crucial modeling choice is that transcription is affine in the regulator abundances. So the transcription rate for a gene equals a basal rate, plus a known perturbation effect from the experimental condition, plus one row of the matrix A multiplied by the regulator vector. That matrix A is the unknown signed regulatory network.

RAMAN: I like that the unknown enters linearly at the transcription step. That linearity is what later lets them convert the entire dynamical problem into sparse linear regression once they integrate the dynamics.

ADA: The observation model gets equal care. Because cells are destroyed when measured, each observation is a snapshot with an unknown latent biological time and a condition label. Those latent times are a genuine source of uncertainty since they must be estimated from the data.

EMMY: And perturbations are modeled as operators that modify regulator activities before they enter the transcription function. The theory deliberately stays agnostic about the mechanism — knockout, knockdown, overexpression, or pharmacological inhibition all fit the same framework.

RAMAN: The forward operator maps parameters to whole trajectories, so the inverse problem is nonlinear in a subtle way. The parameters determine the trajectory before any observation can be generated, which makes this quite different from fitting a curve to fixed points.

ADA: They close the section by listing what makes this harder than classical parameter estimation. You only see snapshots, you don't observe time directly, and you have multiple perturbation conditions producing different trajectories of one underlying system.

EMMY: With that setup in place, the stage is set for the impossibility result. If you only have the control condition, something has to give, and the next page proves exactly what fails.

Page 5 of the paper: ADA: The impossibility result starts with a deliberately simple observation. Under the control condition, the regulator trajectory is a curve through regulator space, and for each target gene the transcription forcing is an affine function of that curve.

EMMY: So every target gene only ever reveals the values of its affine function along that one curve. Whatever the function does away from the curve is completely invisible to the measurement process.

RAMAN: And the geometry takes over from there. They form an augmented feature vector with a constant one on top and the regulator trajectory below. Two parameter vectors are observationally equivalent exactly when their difference is orthogonal to every point on that augmented trajectory.

ADA: Theorem 3 point 2 makes the failure precise. If the span of those augmented feature vectors has dimension less than K plus one, then there is a nonzero parameter perturbation orthogonal to the whole span. That perturbation changes the network but leaves the transcription forcing untouched at every time point.

EMMY: The proof becomes almost trivial once you see it. You pick a nonzero vector in the orthogonal complement of the feature space, and you have constructed two distinct networks that generate identical control trajectories.

RAMAN: And they emphasize the result is constructive and independent of the estimator. No optimization algorithm, no regularization scheme, no statistical inference method can distinguish the two networks from control data alone.

ADA: The dimension condition drives the whole argument. One developmental trajectory typically explores a feature space far smaller than the K plus one dimensions needed to pin down a full row of the regulatory matrix.

EMMY: So the inverse problem is genuinely ill-posed without perturbations. The natural next question is what additional experiments restore uniqueness, and that's exactly what the following section takes up.

Page 7 of the paper: ADA: Section four brings in the perturbations, and the structure mirrors the nonidentifiability argument beautifully. Each experimental condition q produces its own augmented feature vector, now with the perturbation operator applied to the regulator trajectory.

EMMY: So instead of one feature space, you get several, and the question becomes whether their combined span fills the whole K plus one dimensional parameter space.

RAMAN: They package that information into a population information matrix, which is exactly a Gram matrix. Its quadratic form sums the squared inner products of any candidate parameter direction with all the feature vectors across conditions and times.

ADA: Then Theorem 4 point 1 delivers a complete characterization. The regulatory parameters are uniquely determined by the collection of perturbation trajectories if and only if that information matrix is positive definite. Equivalently, if and only if the aggregated feature space spans the full parameter space.

EMMY: The proof splits into two directions, and each one is clean. If the matrix is positive definite and two parameter vectors produce identical forcing everywhere, then their difference multiplied by the matrix gives zero, so the difference itself must be zero.

RAMAN: And the converse is just as elegant. If the matrix is singular, take a nonzero vector in its null space. Since the quadratic form is a sum of squares with positive weights, that vector is orthogonal to every single feature vector, which constructs an indistinguishable parameter perturbation.

ADA: A necessary and sufficient condition like this is rare in inverse problems. Identifiability becomes a spectral property of a matrix you can compute from the experimental design, rather than a vague hope about having enough data.

EMMY: But the whole proof leans on the linear structure of transcription. I'm already wondering what happens when the transcription function is nonlinear, which is precisely where the next page goes.

Page 9 of the paper: ADA: The next page relaxes the linear assumption. For a general nonlinear transcription function, global identifiability is too much to hope for, so the paper aims for local identifiability in a neighborhood of the true parameter.

EMMY: The tool is sensitivity analysis. You differentiate the transcription forcing with respect to the unknown parameters and collect those sensitivity vectors into a continuous-time Gram matrix, in direct analogy with the linear information matrix.

RAMAN: Proposition 4 point 2 states that if that sensitivity information matrix is positive definite, the parameter vector is locally identifiable. The proof is one line: positive definiteness gives the Jacobian full rank, and the inverse function theorem does the rest.

ADA: It's worth being precise about what local means here. You're guaranteed a neighborhood around the truth where no other parameter vector generates the same forcing. That's weaker than the global statement from the linear case, but it's the natural notion for nonlinear models.

EMMY: Then the paper shifts from exact identifiability to practical identifiability, and I think that's the more useful concept for people running experiments. A matrix can be positive definite and still nearly singular, which means certain parameter directions are only weakly excited.

RAMAN: The condition number captures that — the largest eigenvalue divided by the smallest. A large condition number means noise gets amplified along the weakly excited directions, and the finite-sample bounds later scale with the inverse of the smallest eigenvalue.

ADA: So the eigenvalue structure of the information matrix becomes the bridge between structural identifiability and actual predictive performance under held-out perturbations. The same object tells you whether recovery is theoretically possible and how stable it will be in practice.

EMMY: That's a reassuring link between the abstract theory and the noisy reality of measurements.

RAMAN: But everything so far assumes clean trajectories. The next part of the paper has to confront the fact that you only observe noisy snapshots, which calls for a completely different estimation strategy.

Page 11 of the paper: ADA: The estimation strategy opens with a warning. Recovering the network directly from the differential equation would require estimating the time derivative of unspliced RNA, and numerical differentiation amplifies measurement noise and latent-time uncertainty. The authors want no part of that.

EMMY: So they integrate the dynamics over an interval instead. The integrated equation is an exact identity, no approximation involved, and the unknown regulatory coefficients remain linear after the integration.

RAMAN: That's the elegant move. Integration smooths noise rather than differentiating it into something worse. Each integration interval produces one linear equation in the parameters, with features built from the integrated regulator trajectory.

ADA: Concretely, the response is the change in unspliced RNA plus a degradation term integrated over the interval, and the feature vector stacks the interval length with the integrated regulator abundances. The parameter vector is the intercept together with the regulatory row.

EMMY: In practice the trajectories are estimated from snapshots, so both the response and the features carry error. The paper handles that by writing everything as an empirical linear model, with an error term that absorbs all the uncertainty sources.

RAMAN: And that error term is central to the statistical analysis because it has structure. It's a sum of five contributions: counting noise from molecular sampling, errors from estimating latent time, kinetic parameter uncertainty, numerical quadrature error, and model misspecification.

ADA: Reducing a dynamical inverse problem to a noisy linear inverse problem is a major step, because it lets the theory borrow the full machinery of sparse regression.

EMMY: But the design matrix itself is estimated, so it comes with a perturbation too. I suspect that's going to be the main technical challenge in the recovery guarantees they prove later.

RAMAN: Before the guarantees, they need to define the sparse estimator and the reconstruction algorithm, and that's what appears on the next page.

Page 13 of the paper: ADA: On this page the empirical linear model becomes a concrete estimator. For each target gene, they solve a penalized least squares problem — essentially the Lasso — but with one important detail. The intercept, representing basal transcription, is not penalized.

EMMY: That makes biological sense. Sparsity applies to regulatory interactions, not to the baseline transcription rate. You want to shrink the network coefficients toward zero, but the housekeeping part has to stay free.

RAMAN: The error decomposition gets spelled out explicitly: counting noise, latent time estimation, kinetic parameters like degradation rates, numerical quadrature, and model misspecification. Five separate sources, and the theory has to control their aggregate effect through a single effective noise condition.

ADA: That effective noise condition bounds the correlation between the design matrix and the error term by half the regularization parameter. It's the natural adaptation of the standard sparse recovery condition to this multi-source setting.

EMMY: The algorithm itself has two stages. First, for each perturbation condition, you reconstruct the latent trajectories and estimate kinetic parameters. Second, gene by gene, you build the design matrix and solve the sparse optimization problem.

RAMAN: Separating trajectory estimation from network inference is a deliberate choice. It lets the analysis distinguish errors coming from state estimation from errors introduced by the inverse problem itself.

ADA: Then the finite-sample theory begins with a design stability result. If the estimated design matrix is close to the exact one, the smallest eigenvalue of the empirical Gram matrix stays bounded away from zero, by a Weyl eigenvalue perturbation argument.

EMMY: So accurate trajectory reconstruction preserves the conditioning of the inverse problem. That's the foundation the main recovery theorem rests on.

RAMAN: And I'm eager to see how the recovery theorem handles the fact that both the response and the features are noisy. That's the part most classical guarantees assume away.

Page 15 of the paper: ADA: The recovery proof starts with a deterministic basic inequality. From the optimality of the Lasso solution, you get a bound involving the prediction error, the regularization terms, and the correlation between the design matrix and the error vector. Everything else in the proof follows from manipulating that inequality.

EMMY: The next step is translating the inequality into a cone condition. Because the true coefficient vector is sparse, the estimation error has to concentrate most of its mass on the true support — the off-support component is bounded by three times the on-support component.

RAMAN: That cone condition is exactly where the restricted eigenvalue assumption enters. The empirical design matrix only needs to be well-conditioned on sparse directions, and the cone condition guarantees the error vector lives inside those directions.

ADA: From there the algebra is standard but worth walking through. The restricted eigenvalue condition gives a lower bound on the prediction error in terms of the support error, and the basic inequality gives an upper bound. Combining them with Cauchy–Schwarz produces the final bound.

EMMY: Theorem 6 point 3 states it cleanly: the estimation error is at most a universal constant times the square root of the sparsity level, divided by the restricted eigenvalue constant, times the regularization parameter.

RAMAN: And the regularization parameter is chosen to balance two contributions. The stochastic term scales like sigma times the square root of log K over M, and the systematic terms — the deltas for latent time, kinetics, quadrature, and model misspecification — enter additively.

ADA: That split is the real content of the theorem. More integration intervals shrink the stochastic term, but the delta terms remain untouched.

EMMY: So we have a guarantee with two parts that behave completely differently under more data. The next page interprets what that means for experimental design.

Page 17 of the paper: ADA: The interpretation section makes the two-part structure explicit. The stochastic term decreases as the number of effective integration intervals grows, but M is not the raw number of measured cells — it's the number of informative trajectory intervals you can extract from the data.

EMMY: So collecting more cells only helps if they actually inform the trajectory. That resonates with me as geometry: you need coverage along the curves, not just a huge cloud of disconnected points.

RAMAN: The systematic terms — latent time uncertainty, kinetic parameter error, quadrature error, model misspecification — cannot be eliminated by sample size. You could sequence a million cells and those errors persist, because they're baked into the quality of the trajectory reconstruction itself.

ADA: The restricted eigenvalue constant carries experimental meaning too. Well-designed perturbations produce diverse regulator trajectories, well-separated features, and a constant that stays bounded away from zero. Poorly designed perturbations give nearly collinear features and unstable recovery even with large samples.

EMMY: So the theory identifies three requirements: sufficient sampling along the trajectories, accurate reconstruction of latent states, and perturbations that actually excite distinct regulatory directions.

RAMAN: And there's a blunt consequence the paper states directly. Increasing sequencing depth alone cannot remove all reconstruction errors in snapshot single-cell regulatory inference. That's a statement experimentalists should hear loudly.

ADA: It also explains the choice of validation strategy. Predicting responses under held-out perturbations is the honest test of whether the recovered network generalizes, and the identifiability diagnostics predict how reliable that generalization will be.

EMMY: That connection closes the loop between the inverse problem theory and the experimental design.

RAMAN: So we've reached the end of the technical content. Time to pull back and look at the whole arc.

Conclusion: ADA: Time to close the book on this one. What stays with me is the logic flow: unperturbed snapshots fail because of a geometric shortfall in the observable feature space, perturbations restore uniqueness exactly when an information matrix becomes positive definite, and the integral estimator converts the whole inverse problem into a sparse regression with explicit finite-sample bounds.

EMMY: And those bounds separate what you can fix by collecting more data from what you can't. Latent time uncertainty and kinetic parameter errors persist regardless of sample size, so trajectory reconstruction quality and perturbation design are not optional extras.

RAMAN: For me the elegance is in the constructiveness. The nonidentifiability is demonstrated by explicit directions in an orthogonal complement, the identifiability is checked by a matrix you can compute from the design, and the recovery guarantee distinguishes the error terms that shrink from the ones that don't.

ADA: There's also a methodological stance worth honoring. Rather than offering another heuristic algorithm, the paper establishes limits, conditions, and guarantees. That's the kind of foundation this field has been missing.

EMMY: And the held-out perturbation validation means the theory doesn't stay abstract. It ties identifiability diagnostics directly to whether an inferred network can predict cellular responses to new perturbations.

RAMAN: I wonder whether the framework could stretch to partially observed regulators or feedback between target genes, but as a first systematic treatment it covers a remarkable amount of ground.

ADA: Then we'll say goodbye to this paper and get ready for the next one. Thanks for listening, and see you soon.

EMMY: See you next time.