acceptodds
Under review as a conference paper at ICLR 2027

Learning Individual Dynamics from Sparse Cross-Sectional Snapshots

Abstract

Forecasting how an individual unit, such as a patient, a community or a machine, will evolve usually requires repeated measurements of that unit, yet many datasets are cross-sectional: many units, each observed once. Sequence models need the per-unit sequences such data lack, while optimal transport and flow matching describe the population rather than the individual. We introduce CADENCE, which forecasts individual continuous-time trajectories from a single observation per unit by anchoring latent dynamics to static context. A score-based encoder maps observations into a latent space with a fixed Gaussian marginal, and a context-routed mixture of experts blends a shared basis of dynamical archetypes into the per-unit parameters of a latent neural ODE, trained by matching per-subgroup marginals across units. We show that marginal matching alone cannot identify individual dynamics, and give structural conditions under which the per-subgroup dynamics are identifiable within a fixed encoder frame; these conditions motivate the design. On six synthetic benchmarks, CADENCE attains the lowest individual-level error among published cross-sectional methods on five and the highest per-unit skill on all six, while conditional transport trained on full trajectories remains more accurate where such data exist. On real-world lineage-traced haematopoiesis data, trained on day-2 cells and unlabelled later-day populations without lineage or fate labels, CADENCE achieves strongest clonal fate predictions with AUROC , above published unsupervised dynamics methods (-).

Primary area:
learning on time series and dynamical systems
Keywords:
dynamical systems, trajectory inference, neural ODEs, cross-sectional data, score-based generative models, mixture of experts, identifiability

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_wK3HOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes CADENCE, a framework for learning individual continuous-time dynamics from sparse cross-sectional observations, where each individual is observed only once. The method first uses a score-based encoder to map observations into a Gaussian latent space, and then employs a context-conditioned mixture-of-experts to parameterize individual neural ODE dynamics. Training is performed by matching subgroup-level marginal distributions across time without longitudinally paired observations. The authors also provide an identifiability analysis under structural assumptions and evaluate the method on six synthetic or semi-synthetic benchmarks and a real-world single-cell lineage-tracing dataset, reporting improvements over several existing methods.

Critical strengths

None

Critical weaknesses
  1. Line 93. Can the proposed encoder establish a continuous bijection between spaces of different dimensions? What mathematical theory justifies such a construction? I also noticed that in some experimental settings in the appendix, the authors use p=q=500p=q=500. If the encoder does not reduce the dimensionality, what exactly is the motivation for introducing it?

  2. Lines 161–163. The authors begin discussing what appear to be empirical findings before even reaching the experimental section. This is a rather unusual choice of presentation and feels out of place at this point in the paper.

  3. The overall writing is highly disorganized and difficult to follow. The manuscript is overloaded with abbreviations and extensive cross-references to distant sections and appendices. For example, to understand Assumption 2 in Section 3, I had to consult Section 5, Appendix B.3, and Theorem 1 just to piece together what the assumption actually means. Similar cross-references are scattered throughout the entire manuscript. The authors seem to have distributed closely related material across different parts of the paper without a clear organizational logic, making the manuscript extremely difficult to read. I found it nearly impossible to develop a coherent understanding of the work by reading it sequentially. In my opinion, the paper requires a substantial and careful restructuring and rewriting before it can become a useful and accessible research contribution to the community.

  4. The mathematical presentation appears unnecessarily and deliberately complicated. This is particularly evident in Section 3. Could the authors explain, in plain language, what this section is actually trying to establish? I have received formal training in mathematics, yet I still find the presentation highly confusing. The extensive formalism and terminology seem to obscure the underlying ideas rather than clarify them.

  5. The logical flow of the manuscript is also problematic. One concrete example is Lines 359–366, where the authors introduce a paragraph entitled "Where the advantage comes from." The subsequent discussion lists several experimental results, but seems to lack the essential explanatory statements that would actually clarify where the claimed advantage originates and why these results support that conclusion. This is representative of a broader problem throughout the manuscript: numerical observations and technical statements are presented without sufficient logical connections or interpretation, leaving the reader to infer the intended conclusions.

  6. Line 222. The statement "Assume DFA, COA′, BAS, POS, ENR, PRW and FOA, plus SEP under COA′(b)" is a particularly striking example of the poor presentation. This style of writing is extremely difficult to follow. Requiring readers to repeatedly look up a long list of abbreviated assumptions, some of which have multiple variants and are defined in different parts of the paper, creates an unnecessary burden and seriously undermines readability.

  7. The presentation of experimental results in Section 8 is particularly unsatisfactory. In my view, this section consists largely of an overwhelming enumeration of numerical results, with little prioritization of the important findings and insufficient intuitive explanations or meaningful insights. Rather than guiding the reader toward an understanding of what has been learned from the experiments, the authors simply present one numerical comparison after another. As a result, the section reads more like an automatically generated compilation of experimental results than a carefully structured scientific analysis. From a reader's perspective, it is very difficult to extract useful scientific insights from such a presentation. I believe the manuscript requires a serious and comprehensive reorganization and rewriting, rather than merely minor language polishing.

Questions for authors

I have no specific question for authors. My major concerns are the clarity and readability of theis paper. See weakness.

1 comment

No rebuttal has been received, therefore, I maintain my initial assesement.

Reviewer_yRe6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

CADENCE addresses individual trajectory forecasting from purely cross-sectional data, where each unit is observed only once together with its observation time and static context. The method first maps observations through a score-based probability-flow ODE encoder into a Gaussian-pinned latent space; on BM1–BM6 this encoder is context-indexed. A context router then produces a convex mixture of expert parameters, which FiLM-condition a latent neural ODE, and the model is trained by matching routing-weighted predicted and kernel-smoothed reference distributions across time and subgroups rather than by observing longitudinal trajectories of the same units.

The theoretical analysis first shows that marginal observations alone do not identify individual dynamics, then studies what additional structural assumptions can recover subgroup-specific dynamics within a fixed latent frame. In particular, the paper characterizes the residual ambiguity induced by Gaussianizing encoders, gives parameter-identifiability results under assumptions on subgroup structure, context observability, enrolment and flow observability, and derives a Liouville-based sufficient condition for the latter. The method is evaluated on six synthetic or semi-synthetic benchmarks with simulated longitudinal ground truth and on the real LARRY lineage-tracing dataset.

Critical strengths

The paper addresses a relevant problem: forecasting individual trajectories when each unit is observed only once. The pooling mechanism is intuitive, since later observations from units with similar context can constrain the future of earlier units, and the proposed architecture provides a concrete way to implement that idea.

The evaluation spans seven benchmarks and includes extensive ablations, matched-representation controls and checks of individual forecasting accuracy. The authors explicitly discuss several differences between theory and implementation, report statistical qualifications, and use evaluation splits by clone for the real-data experiment.

Critical weaknesses

W1. Theory: The theoretical guarantees rely on strong assumptions that are difficult to verify in practice, and it is not always clear which parts of the deployed model are actually covered by those guarantees.

W2. Comparisons: A substantial part of the performance gain, especially on BM1, appears to come from the learned representation itself; when competing methods are evaluated in the same latent space, the gap becomes much smaller.

W3. Statistics: Several of the reported advantages over competing methods do not remain significant after correction for multiple comparisons, so some of the superiority claims should be stated more cautiously.

W4. BM7: On the real-data benchmark, CADENCE has access to unlabelled later-time populations that the supervised classifiers do not see, so that comparison does not isolate the benefit of the dynamics model itself.

Presentation: The paper is extremely dense, introduces many acronyms and assumptions very quickly, and often postpones essential explanations or qualifications, which makes the main argument difficult to follow.

Questions for authors

For CVFM/COT-FM in Tab. 2, was the source noise fixed per unit across query times, and how do increment ratio change if it is?

What do the cross-sectional baselines reach on the unconditional encoder's latents?

1 comment

As the discussion period ends I have received no rebuttal, therefore kept my score

Reviewer_RGGuOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper deals with a very important question: Forecasting how an individual "unit" (called this way in the paper but could be a patient, or a machine) will evolve requires repeated measurements over time. This are called longitudinal datasets. However, many real-world datasets are cross-sectional (only a single observation per patient). Traditional sequence models fail due to the lack of histories, while optimal transport or flow matching methods describe the aggregate population rather than the individual. The authors propose a solution to be able to predict at the individual level without longitudinal data. The underlying model involves a latent mixed effect ODE model. Experiments have been done on 7 datasets, one on real medical data.

Critical strengths

The theory seems robust (but the paper is too difficult to read to be completely sure of every claim.) Experiments are showing improves on 5 of the 6 benchmark dataset and interesting results on the real medical dataset. The performance gains in Mean Absolute Error (MAE) and sliced 2-Wasserstein distance (SW2) are substantial compared to published baselines. Its ability to accurately predict the fate of real cells without any fate labels during training demonstrates excellent biological applicability. The decoupled training strategy (Stage 1 then Stage 2 offline) avoids end-to-end backpropagation through cascaded spatial and temporal ODEs, significantly limiting training time.

Critical weaknesses

The paper is hard to read and notations are not defined when introduced. In particular Proposition 2 uses notations like # (pull back or push forward?) and refers to theorems that are later in the appendix and are not summarized when used. This same critics holds for Assumptions 1 and the followings. This makes the paper way too difficult to follow. The framework relies on strong assumptions regarding the data-generating process. As the authors note, these structural properties cannot be formally verified from a trained model alone. The method only outputs point forecasts. The lack of uncertainty quantification currently limits its deployment in safety-critical scenarios, such as clinical or medical decision-making.

Questions for authors

Can you re-write the paper to make it easy to follow and introduce notions and results before they are used? Can you give a way to compute confidence intervals maybe by introducing a SDE instead of an ODE?

1 comment
Reviewer_RGGuOpenReview ↗

No rebuttal

I keep my score.

Reviewer_3PEoOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes an algorithm for inferring the underlying dynamics of a dynamical system in a latent space. The algorithm consists of several components: an encoder-decoder component, a context conditioning component and a neural ODE component. The encoding into the latent space is done with the help of an ODE flow model, which is trained separately from the main algorithm. Context is encoded with the help of a separate context model. The training procedure (as per Algorithm 1 in Appendix C1) is done by matching the marginals of simulated dynamics, obtained by integrating a learnt drift model, with the true distributions at each time step. The algorithm is tested on a wide range of benchmarks.

Critical strengths

The paper contains promising results on both synthetic and real-world benchmarks. However, it is quite difficult to fully appreciate the results without a complete theoretical justification of the algorithm, which, in my opinion, the paper falls short of delivering.

Critical weaknesses
  1. Unfortunately, the presentation of the algorithm is difficult to follow. The following parts are confusing:
  • The main algorithm is presented only in the appendix; some losses used in the algorithm are not defined (e.g., LmassL_\text{mass}), and some are difficult to understand, e.g., LfeatL_\text{feat} uses yiy_i and y^i\hat{y}_i, but they are not defined in the pseudocode. Moreover, I would strongly suggest making each loss function explicitly dependent on the parameters it is defined over: L(θ)=some expression depending on θL(\theta) = \text{some expression depending on } \theta.
  • The components of the main training objective (line 200) are not defined in the main text.
  • Proposition 2 in Section 2 references Theorems 1 and 2, although the theorems themselves are only stated in Section 4. Moreover, the proposition contains references to baselines in its statement, which makes me wonder whether the proposition applies only to a fraction of application cases rather than all of them.
  • There are at least 5 assumptions, and BAS and ENR are not highlighted as separate assumptions, although they are referenced as such in Theorem 1. Besides, it is difficult to fully appreciate the importance of these assumptions, as no explanation is given for any of them. Nor do the authors explain why these assumptions are necessary.
  1. Some parts of the text are overloaded with terms that are not properly defined or are simply difficult to understand:
  • Abstract, line 17: a context-routed mixture of experts blends a shared basis of dynamical archetypes into the per-unit parameters of a latent neural ODE, trained by matching per-subgroup marginals across units, where phrases like dynamical archetypes and context-routed mixture of experts blends are not properly explained.
  • Line 51: ... a late-stage unit's snapshot constrains the future of a contextually similar early-stage unit, ...
  • Line 78: ... that late-stage snapshots supervise contextually similar early-stage units without any pairing.
  • Lines 124–126: the terminology is difficult to understand without a schematic or an intuitive explanation.
  1. I find that the statements of the theorems and assumptions are not rigorous. Specifically:
  • I struggle to see what the following statement means: Under reading (a) the context supplies the leaf partition; under reading (b) FOA and SEP recover it. I am left guessing whether readings (a) and (b) refer to Assumption 2 or something else. Moreover, I do not understand what is meant by the context supplies the leaf partition and FOA and SEP recover it (what is the leaf partition?).

As a general comment, I find the presentation of the theory quite messy and confusing. Many concepts are left unexplained, and Sections 3 and 5 are separated by the implementation details given in Section 4, which do not seem directly related to the conceptual presentation of the paper. All in all, despite carefully reading Sections 1 to 5, as well as parts of the appendix, I remain perplexed about what precisely the paper proposes. On these grounds, I venture to suggest making critical amendments to the key sections of the paper.

Questions for authors
  1. Why is it necessary to encode the data in the latent space? Why would the proposed method not work directly in the data space?

  2. The encoder seems to map data points to Gaussian samples, but the dynamics in the latent space then somehow move the points. Why are these points guaranteed to remain distributed according to the Gaussian distribution, which is assumed by the decoder?

  3. What motivates the choice of a flow model as an encoder? Can a VAE be used as well?

  4. What makes the proposed method critically differ from, for example, OT-CFM (Tong et al., 2024) or MFM (Kapuśniak et al., 2024)? Can the context model be added directly to these methods?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.