acceptodds
Under review as a conference paper at ICLR 2027

TRACE: Transcriptional Regulator Modulated Reference Geometry for Temporal Single-Cell Flow Matching

11 views

Abstract

Progenitor-exhausted (Tpex) CD8 T cells retain self-renewal and proliferative capacity, making their differentiation therapeutically important. Resolving this progression from temporal single-cell RNA sequencing is difficult because destructive sampling provides population snapshots without directly observed cell trajectories. Here, we introduce **TRACE**, a regulatory-informed conditional flow-matching model for predicting unobserved temporal populations. To incorporate regulatory direction into transport, we map temporal derivatives of expression-derived transcription-factor regulon activities into latent-space tangents for shape-preserving cubic-Hermite paths between optimal-transport-coupled populations. Across three T-cell exhaustion studies, the clearest improvements occurred at earlier holdouts. On the Gupta day 7 development holdout, was 21.7% lower than with linear CFM and 8.8% lower than with cubic CFM. Improvements extended to the earlier Giles and DSM holdouts, including a reduction in DSM day 8 from with linear CFM to with TRACE. At later holdouts, linear or alternative transport often matched or exceeded TRACE performance. These results support stage-dependent benefits of regulatory-informed reference paths for temporal interpolation, with the largest gains during earlier exhaustion-associated transitions.

Primary area:
learning on time series and dynamical systems
Keywords:
single-cell RNA sequencing, T-cell exhaustion, population dynamics, optimal transport, context-aware, conditional flow matching, temporal interpolation

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

This paper studies the reconstruction of cell population dynamics from single-cell gene expression snapshots collected at multiple time points, with a focus on predicting cell-state distributions at unobserved intermediate times. The method uses optimal-transport-based conditional flow matching: static OT couplings between consecutive observed time points define temporal cell tuples, from which reference paths and their velocity targets are constructed. Its main contribution is to incorporate an external TF–target regulatory prior to compute cell-level regulon activities, map their temporal changes into PCA space, and use the resulting directions to modulate the tangents of shape-preserving cubic-Hermite reference paths. The authors evaluate the method through temporal holdout experiments on three T-cell exhaustion datasets, alongside controls that permute cell–activity correspondence or TF activity channels. The results show lower mean prediction errors at earlier holdouts, whereas linear CFM or alternative transport methods often perform better at later holdouts.

Critical strengths
  1. The paper introduces an interesting way to guide flow matching reference paths using transcription-factor regulatory information. It maps temporal changes in regulon activity into latent-space tangents. This is a meaningful contribution because temporal snapshots alone do not uniquely determine the intermediate dynamics, and regulatory information offers an additional basis for choosing reference paths.
Critical weaknesses
  1. The benefit of regulatory information remains unclear. TRACE maps the activities of 29 regulons into a 30-dimensional PCA space using ridge regression and uses the mapped activity derivatives to modify reference-path tangents. However, the PCA representation may capture gene expression changes that are not adequately represented by these regulons, and the paper does not quantify how much latent-state variation the activity representation explains. The directional alignment analysis in Table 9 provides some supporting evidence but does not resolve this concern. Furthermore, on Giles day 15 in Table 7, both permutation controls outperform TRACE in all five runs, indicating that preserving the correct regulatory correspondence does not provide an additional predictive benefit in this setting.

  2. The predictive benefits have limited scope. TRACE achieves lower mean errors at earlier holdouts, but its mean W2W_2 does not improve over linear CFM at any of the three later holdouts and is consistently worse than moscot. Moreover, the largest improvement occurs on Gupta day 7, which was used to select the steering coefficient, whereas the improvement over cubic CFM on the external Giles day-8 holdout is small and inconsistent across runs. Although the authors acknowledge this stage dependence, the analysis does not adequately explain which properties of the data or dynamics determine whether regulatory guidance is beneficial, nor does it provide a way to assess its suitability using retained data. Consequently, the practical advantage of the additional regulatory information and reference-path construction remains limited.

Questions for authors
  1. Should regulatory information also inform the static OT coupling? The current method determines cross-time cell correspondences using only squared Euclidean distances in PCA space, then uses regulatory information to adjust the reference paths connecting those cells. Given that regulatory information is intended to guide the dynamics, why do the authors use it only to modify paths rather than also to inform the coupling? Could matching cells solely by expression distance connect cells with incompatible regulatory changes, thereby limiting the effectiveness of subsequent tangent adjustments?

  2. How well do the selected regulons represent changes in the PCA state space? Please clarify the overlap between the target genes used for regulon activity scoring and the genes used to construct the PCA representation. If these gene sets differ, how do the authors justify using changes in the selected regulon activities to guide dynamics in a PCA space that may also capture variation in genes outside those regulons? The directional alignment analysis in Table 9 provides some supporting evidence but does not establish how much of the overall latent-state variation these regulatory features explain.

  3. Does the regulatory information provide an additional predictive benefit? On Giles day 15 in Table 7, both permutation controls outperform TRACE in all five runs. Could the authors explain why disrupting cell–activity correspondence or TF-channel alignment improves prediction in this setting? What evidence distinguishes the contribution of authentic regulatory information from the effects of feature aggregation and tangent modification?

Reviewer_spWLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors introduce a method they call TRACE, where flow-matching cubic paths are reshaped with expression-derived transcription-factor regulon activities. In the experiments, they show that it improves reconstruction of heldout snapshots in three studies (in early snapshots), but simpler methods perform better at later snapshots.

Critical strengths

The paper reports both failure modes and gains (the conclusion from the experiments is that TRACE wins on early heldout snapshots but loses to linear CFM or moscot on later ones).

Critical weaknesses

The biological source of gains is not super clear. Indeed, the regulon-permutation controls do not test if the signed TF–target prior adds information beyond a feature map derived from expression with equal size. The benefits of the approach are not big. For example, on the Giles day-8 experiment, the gain in W2 over cubic CFM is only 1.6% with non negligible variance, and on Giles day 15, the method is much worse than linear CFM (see Tab 3). It is not clear how one should decide on whether or not to use the method.

Questions for authors

If you refit the ridge map, how does TRACE compare with corresponding random TF–target priors or expression-derived features of equal dimension? Can you provide a rule for how to select the steering strength or when to not use steering? And can you show that this rule correlates with the regimes shown in the experiment (early vs late heldout snapshots)?

Reviewer_2Ae5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces TRACE, a regulatory-informed conditional flow-matching model for predicting unobserved temporal populations in temporal scRNA-seq. TRACE uses optimal transport to link cells across observed time points and constructs shape-preserving cubic-Hermite reference paths through these linked states. Its main contribution is to steer the reference-path tangents using temporal derivatives of transcription-factor regulon activities, mapped into PCA space through a fitted ridge regression. This introduces regulatory structure into the flow-matching supervision while retaining the same OT-linked cell states. TRACE is evaluated on three T-cell exhaustion datasets with intermediate developmental time points held out. TRACE achieves lower mean prediction error than linear and cubic CFM at the earlier holdout in each study. At later holdouts, however, linear CFM or alternative transport methods often match or outperform TRACE.

Critical strengths
  1. Using TF-regulon activity changes to guide reference-path directions is a natural and interpretable approach to incorporating biological structure into flow matching. TRACE modifies the reference geometry that defines the training targets. Importantly, the steering coefficient ρ\rho is selected on Gupta day 7 and then kept fixed for the Giles and DSM evaluations, where early-stage improvements are also observed. These results suggest that the predictive benefit is not restricted to the dataset used for tuning.
Critical weaknesses
  1. Both the regulon activities and the PCA coordinates are derived from the same expression measurements. TRACE transforms expression into regulon activities and then learns a ridge map from those activities back into PCA space. Although the TF–target prior introduces external biological structure, the observed improvements do not establish whether the correctness of those relationships is important. A plausible alternative explanation is that the activity representation provides useful smoothing or dimensionality reduction, independently of its specific regulatory interpretation. The existing permutation controls do not fully resolve this issue. TF-channel permutation retains the ridge map fitted to the original activities, disrupting the correspondence between the features and their fitted mapping rather than comparing two independently fitted representations. The authors acknowledge that these controls do not establish the superiority of the authentic regulatory graph over a matched random representation. A more convincing control would construct randomized or degree-/size-matched TF–target priors, recompute the activities, and refit the activity-to-latent map for each control under otherwise matched conditions. This is consequential to my assessment because the use of regulatory knowledge is central to the contribution. I find the predictive improvements more convincing than the evidence that biologically meaningful TF–target relationships specifically account for them.

  2. The benefits are stage dependent, and sensitivity to the steering coefficient remains insufficiently characterized. With \rho=1.0 selected on Gupta day 7, TRACE improves prediction at earlier holdouts but is often matched or outperformed by linear CFM or moscot at later holdouts. Although using a fixed coefficient provides a useful test of transfer, the reported sweep is limited to Gupta day 7, leaving unclear whether alternative steering strengths could mitigate the later-stage underperformance. The current method also provides no demonstrated criterion, based only on retained populations, for identifying when regulatory steering is beneficial. This limits the evidence for its practical applicability across developmental stages.

Questions for authors

Clarification questions: Can the authors explain what is resampled across the five matched runs: retained-cell subsets, biological samples, network initialization, OT mini-batches, or evaluation subsets? If training populations differ between runs, how are those populations constructed?

Comments on the paper

2 comments
Area_Chair_vRbmOpenReview ↗

Please Join the Discussion

Dear Reviewers,

The authors have submitted their responses. Please read them and provide your feedback. Thank you!

Best regards, Area Chair

AuthorsOpenReview ↗

Global Response

We thank the reviewers and Area Chair for the thoughtful discussion. Reviewer 2Ae5 has now confirmed that we addressed the concerns raised in the review. During rebuttal, we expanded the benchmark to 30 matched seeds, completed independently refitted controls, tested alternative steering schedules, and evaluated reference selection on a fourth, independent time course. We appreciate the recognition from Reviewers 2Ae5 and bxTS that regulatory activity offers a meaningful way to guide flow-matching reference paths.

Importance of reconstructing early T-cell differentiation

Single-cell RNA sequencing measures different cells at each time. As such, reconstructing an unobserved population requires assumptions about how cells change between measurements. With TRACE, we use changes in transcription-factor regulon activity to guide the paths used to train conditional flow matching. We retain the same observed populations and optimal-transport correspondences, allowing us to test the contribution of regulatory guidance directly. Across three studies, we improved reconstruction at the early holdouts while keeping the steering coefficient fixed after selecting it on Gupta day 7.

New experiments for each reviewer

  • Reviewer 2Ae5 asked whether the gains depend on authentic regulatory connections and what varies across runs. We independently recomputed features and refitted the models for degree-preserving network rewires, random gene sets, and expression controls of the same dimension. Reconstruction accuracy was comparable across authentic and refitted rewired networks. To test the biological information in the connections, we then evaluated exhaustion-state classification on independent Daniel data. Authentic regulatory features achieved 52.4% macro recall versus 45.0% across 100 rewired networks, ranking third among all 101 networks. We also tested the importance of trajectory-specific guidance. Shuffling mapped directions increased Giles day-8 reconstruction error from 18.097 to 18.674. For reproducibility, we specified that each seed resampled 80% of retained cells, with matched populations and training conditions across methods.

  • Reviewer spWL asked whether the early improvements were reproducible and how to choose the steering strength. We repeated the external benchmark across 30 matched seeds, selecting variable genes exclusively from observed populations. At Giles day 8, mean reconstruction error, measured by W2W_2, fell from 21.122 with linear CFM and 19.010 with cubic CFM to 18.499 with TRACE. We improved over cubic CFM in 28 of 30 runs. At DSM day 8, we reduced error from 11.098 and 11.608 to 10.585, improving over both baselines in all 30 runs. In separate steering sweeps, the coefficient selected on Gupta also achieved the lowest mean error within the tested cubic-Hermite family at all four external holdouts. These repetitions measured computational and cell-resampling variability.

  • Reviewer bxTS asked how well the regulatory representation captures cell states, whether regulation should also inform cell matching, and when to choose TRACE on a new dataset. Authentic regulons explained more cross-validated PCA variation than refitted rewires at all six holdouts. Adding regulatory activity to the transport cost changed 47.3% of diagnostic cell correspondences and modestly improved reconstruction, with TRACE outperforming cubic CFM under both matching procedures. We also tested constant blends and four time-dependent schedules. The best constant blend matched or exceeded every time-dependent schedule. Because linear CFM performed better at later holdouts, we developed two selection rules based on source position and tangent speed, using only observed populations. On the independent Daniel day-8 holdout, both rules selected TRACE before we evaluated the missing population. We then achieved 0.410 lower mean W2W_2 than linear CFM, with a 95% computational/resampling interval of [0.347, 0.473] and improvements in all five matched runs.

Case for acceptance

Through these additional experiments, we reproduced the early reconstruction gains, tested the regulatory representation against the requested controls, and successfully selected TRACE on an independent time course using information available before prediction. We believe this combination of a biologically motivated change to flow-matching supervision and reproducible improvements across studies merits acceptance. We also connected reconstruction to named regulatory programs and demonstrated the value of authentic connectivity for exhaustion-state classification. We are grateful for the questions that prompted this work. With Reviewer 2Ae5 now confirming that we addressed the concerns raised, we respectfully ask the Area Chair and reviewers to consider the expanded evidence, reconsider the scores, and support acceptance of TRACE.

Thank you,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.