acceptodds
Withdrawn from ICLR 2027

Mapping single cell differentiation fates with FateODE

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

15 views

Abstract

Single-cell transcriptomics provides a powerful lens for deciphering cellular differentiation dynamics and understanding how phenotypic diversity emerges. However, because sequencing is inherently destructive, temporal progression cannot be tracked longitudinally within individual physical cells. Existing paradigms therefore diverge into two separate regimes: recovering macroscopic differentiation trends via trajectory inference, or approximating instantaneous dynamics through RNA velocity. Fundamentally, velocity and trajectory are the differential (local derivative) and integral (global path) manifestations of the same underlying dynamical system. While recent continuous dynamical frameworks (such as Neural ODEs) attempt to unify these perspectives, they exhibit three critical deficiencies: 1) failing to strictly enforce bidirectional consistency between velocity fields and trajectory integrals; 2) insufficiently accommodating technical measurement noise inherent to sequencing; and 3) inaccurate depiction of intrinsic stochasticity in cell fate bifurcations. Consequently, inferred dynamics frequently drift along biologically implausible off-manifold paths or overfit local noise, yielding erratic and overly tortuous trajectories. To overcome these limitations, we propose , a stochastic Neural ODE framework that enforces rigorous derivative-integral consistency while capturing manifold-constrained stochastic fate transitions. FateODE harmonizes macroscopic trajectories with splicing velocities by anchoring continuous integration across discrete waypoint observations and regularizing point-wise derivatives with high-confidence velocity vectors. To capture bifurcation stochasticity without sacrificing manifold geometry, FateODE introduces an inference-time diffusion term parameterized by a decoder-only variational architecture under Maximum Mean Discrepancy (MMD) supervision, faithfully simulating Brownian motion directly on the transcriptomic manifold. Beyond reconstructing smooth developmental vector fields, FateODE functions as a continuous "world model" capable of projecting cellular fates, prioritizing differentiation driver genes, and forecasting fate-landscape shifts under in-silico genetic perturbations. Extensive benchmarks across synthetic and real lineage-traced datasets demonstrate FateODE's superior dynamic modeling fidelity, accurately predicting terminal fate distributions under both unperturbed and perturbed conditions. https://anonymous.4open.science/r/FateODE-83D8/

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single-cell transcriptomics, transcriptome dynamics, Neural Ordinary Differential Equation, cell differentiation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_QxNhOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces FateODE, a stochastic Neural ODE framework for learning single-cell differentiation dynamics from snapshot data. The method enforces derivative–integral consistency between trajectory integrals and splicing velocities through waypoint-anchored path integration, bidirectional supervision, and velocity regularization on a quality-filtered gene set (SQS), and adds an inference-time diffusion term parameterized by a decoder-only variational architecture under MMD supervision. The method is evaluated on hematopoietic lineage-tracing data (Li et al., 2023), a pharyngeal endoderm dataset with a held-out developmental time point, and in silico perturbation analyses of hematopoietic fate territories.

Critical strengths

The problem of reconciling RNA velocity with trajectory inference is real and important, the manuscript is clearly written, and an anonymous code repository is provided.

Critical weaknesses

Critical Weaknesses

  1. Substantial overlap with VeloTrace. VeloTrace ("Reconciles Divergent Velocity and Trajectory in Single-cell Transcriptomics with Deep Neural ODE," 2026) shares substantial methodological similarities with this work, including the Splicing Quality Score (SQS), velocity alignment, and multi-scale temporal sampling. Despite these overlaps, the manuscript does not cite or discuss VeloTrace, nor does it provide a comparison. The authors should clarify the methodological novelty of FateODE relative to this closely related work.

  2. Overstated claims in the abstract. The abstract claims "Extensive benchmarks across synthetic and real lineage-traced datasets," yet no synthetic experiments are presented in the main text. It also claims accurate fate prediction under perturbed conditions without independent experimental validation. Furthermore, the claim of "faithfully simulating Brownian motion directly on the transcriptomic manifold" is insufficiently justified, as the implementation relies on neighborhood-based MMD matching in VAE latent space without a formal characterization of the corresponding stochastic process.

  3. Insufficient ablation studies. Although Figure 4 presents a limited component substitution experiment, the manuscript lacks systematic ablations of multi-scale waypoint supervision, bidirectional reconstruction, SQS-based velocity regularization, and manifold-constrained diffusion. Given the overlap with VeloTrace, it remains unclear whether the improvements primarily originate from previously established components or the newly proposed stochastic module.

  4. Lack of perturbation validation. The in-silico perturbation experiment in Section 4.3 identifies candidate regulators through correlations with model-predicted fate probabilities and then simulates their joint upregulation using the same model. However, no independent perturbation ground truth is provided to validate the predicted fate shifts. Qualitative agreement with existing biological knowledge is insufficient to establish reliable perturbation prediction.

  5. Unclear biological interpretation of stochastic dynamics. The stochastic module learns to match local distributions of neighboring cells through MMD, but these distributions do not necessarily represent actual stochastic differentiation dynamics. Local cell densities are influenced by proliferation, death, and sampling biases. It remains unclear whether the model captures genuine fate bifurcations or simply reproduces the geometry of the observed cell distribution.

  6. Insufficient justification of derivative-integral consistency. The RNA velocity supervision uses cosine similarity, which constrains velocity directions but not magnitudes. Meanwhile, trajectory integration relies on inferred pseudotime, whose scale may not correspond to the temporal scale of RNA velocity. The manuscript does not adequately explain how these two supervisory signals are reconciled to produce quantitatively meaningful dynamics.

  7. Limited quantitative evaluation and statistical reliability. The evidence for manifold adherence and gene-level velocity accuracy in Figure 3 is largely qualitative, with only four genes examined. More systematic quantitative evaluations are needed, along with results across multiple random seeds and statistical significance analysis.

Questions for authors

I have no additional questions for the authors. My main concerns are outlined in the Critical Weaknesses section, and I believe they substantially weaken the methodological novelty and the evidence supporting the paper's central claims.

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

This paper proposes FateODE, a two-stage framework that combines a deterministic Neural ODE in gene-expression space with an inference-time local neighborhood noise generator to model single-cell differentiation dynamics. The deterministic drift field is trained via forward and backward integration losses along sampled pseudotime waypoint sequences together with cosine alignment against filtered RNA splicing velocities, while stochasticity is injected at inference time using a variational generator trained with Maximum Mean Discrepancy (MMD) against local neighborhood distributions. The authors evaluate FateODE on mouse hematopoietic lineage-tracing data (Darlin) and pharyngeal endoderm development (mTEC/BC), and illustrate an unvalidated in-silico four-gene upregulation simulation.

Critical strengths
  1. Combining multi-waypoint forward/backward integration consistency with directional cosine regularization from quality-filtered splicing velocities provides a reasonable heuristic to prevent Neural ODE trajectories from overfitting single-cell measurement noise or drifting far from observed transcriptomic manifolds.

  2. On the hematopoietic clonal dataset and the mTEC/BC E11.5 holdout task, FateODE achieves higher terminal-fate distributional similarity and branch coverage than several recent neural trajectory baselines (scTour, scDiffEq, and DynNet).

Critical weaknesses
  1. The paper suffers from overclaiming—most notably framing a dataset-specific descriptive Neural ODE as a cellular "world model" capable of forecasting genetic perturbations—and contains multiple unsubstantiated claims in the Abstract. Specifically, the Abstract claims extensive benchmarks across "synthetic and real lineage-traced datasets" and accurate fate prediction under "both unperturbed and perturbed conditions," yet the manuscript and two-page supplementary material contain zero synthetic experiments and zero quantitative validation on real genetic perturbation datasets; moreover, because vθ(x)v_\theta(x) is fit purely to observational correlations without causal structure or gene regulatory network priors, shifting four genes in silico cannot establish causal perturbation fidelity.

  2. The deterministic training objective (LfwdL_{fwd} and LbwdL_{bwd}) is mathematically inconsistent with multi-lineage branching and omits essential algorithmic definitions. Because an autonomous ODE dx/dt=vθ(x)dx/dt = v_\theta(x) yields a unique deterministic trajectory from any initial progenitor state xi,0x_{i,0}, minimizing pointwise Euclidean distances against multi-branch waypoints would collapse trajectories toward the conditional branch mean unless the waypoint sequences Si=(xi,k,ti,k)S_i = {(x_{i,k}, t_{i,k})} are pre-segregated by branch; however, neither the main text nor the appendix explains how SiS_i and pseudotimes ti,kt_{i,k} are constructed from unpaired snapshots.

  3. The deterministic drift Φτ\Phi_\tau and stochastic transport map are trained in a completely decoupled manner, and the inference-time update rule X(t+τ;x0)=Φτ(X(t;x0))+gψ(Φ(X;x0),W)X(t + \tau; x_0) = \Phi_\tau(X(t; x_0)) + g_\psi(\Phi(X; x_0), W) lacks a valid stochastic differential equation (SDE) formulation. In particular, omitting the standard Itô time-step scaling τ\sqrt{\tau} on the diffusion term makes the trajectory variance and resulting terminal-fate proportions depend arbitrarily on the discretization step size τ\tau chosen at inference time.

  4. Empirical comparisons omit the most relevant established baselines for vector-field perturbation and fate probability mapping, notably Dynamo (Qiu et al., 2022), which pioneered splicing-based continuous transcriptomic vector fields and in-silico genetic perturbations, as well as CellRank 2 (Weiler et al., 2024) and PRESCIENT (Yeo et al., 2021). Furthermore, the manuscript shows signs of hasty preparation, including notation inconsistencies (gψg_\psi in Section 3 vs. QϕQ_\phi in Section 3.2), a misspelled section header ("NERUAL ODE" in Section 3), three consecutive copy-pasted repetitions of the same sentence in Appendix A.1 (lines 663–667), and the complete absence of hyperparameter tables or multi-seed standard deviations.

Questions for authors

see weaknesses

Reviewer_yXKBOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This submission proposes a new dynamical modeling framework to model single cell differentiation dynamics. Authors call the model stochastic neural ODE that aims to solve challenges that previous neural ODE models have by ensuring consistency between time derivative and integral trajectories by using splicing velocities.

From Introduction, it is difficult to parse the methodological contribution of this work. Methodologically, the proposed model is a type of time-discretised flow that consists of a deterministic component and additive stochastic part. Authors describe that the deterministic part is an ODE flow induced by the velocity field, but specific mechanisms of how this model component is defined exactly are not defined. The stochastic component is driven by white noise, but otherwise the specific mechanisms are not defined.

Regarding model training, authors introduce intermediate points along the trajectory called waypoints. These seem to correspond so-called shooting variables that are used for neural ODE models. The state dynamics start from an initial state and proceed along the intermediate waypoints, following the underlying system dynamics. scRNA-seq data is noisy and hence the use of observed gene expression patterns cannot be used as initial states as such. To account for this noise, authors propose so-called bidirectional supervision that complements forward dynamics with reverse dynamics that should help against the inconsistency introduced by the noisy starting points. The proposed model also claims to regularise the learning task for the underlying system (or vector field) by aligning the forward-simulated trajectories with splicing-derived reference velocities. For some of the details of this regularization, authors refer to Appendix, but the Appendix is missing. Conceptually, such an alignment could make sense, assuming reference velocities could be defined (externally) sufficiently accurately.

Authors propose that the stochastic model component is a transport map that depends on neighbours in its latent features. Description of the method is not precise enough to understand what authors mean and the proposed methodology results in an VAE training loss. Again, for further details authors refer to Appendix, which is missing.

Authors report good performance in predicting terminal-fate proportions and in trajectory reconstruction that are considered in earlier publications.

Overall, while it may be possible to get an understanding of the proposed modeling method, method description does not meet scientific standards.

Critical strengths

None.

Critical weaknesses

Manuscript lacks clarify: introductory sections are difficult to follow and comprehend, and methods description is superficial and lacks lots of details, making it impossible to understand overall and to assess novelty.

Questions for authors

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

Related papers

Discussion (0)

Sign in to comment.