acceptodds
Withdrawn from ICLR 2027

TransFlow: Source-Initialized Flow Transport for Cross-Modality Translation in Single-Cell Multi-Omics

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Single-cell multi-omics provides complementary molecular measurements, but many cells are profiled in only one modality, motivating cross-modality translation of unmeasured molecular states. Conditional generative approaches often use the observed source modality as context while initializing the target-generation trajectory from noise. For paired data, however, the source and target measurements describe the same cell, making the encoded source a natural candidate for the initial state. We propose TransFlow, a source-initialized flow framework that transports paired RNA and ATAC representations between modality-specific latent spaces. Block-wise variational autoencoders encode the two modalities, and a shared bidirectional interval-conditioned velocity field models transport in both directions. A controlled noise-initialized variant receives the same source information as conditioning but begins from Gaussian noise, separating trajectory initialization from source availability. Multiple-run evaluations on BMMC and PBMC10k show that TransFlow obtains the lowest distributional scores in the main comparison, while feature-level results vary across datasets and directions. A single-run MiniAtlas experiment shows a similar distributional trend, and TransFlow obtains the best reported results in the evaluated BMMC held-out-cell-type setting. These findings support source-initialized transport as a useful formulation for distribution-level alignment, while fine-grained feature recovery remains challenging.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single-cell multi-omics, cross-modality translation, RNA-ATAC integration, flow matching, conditional generative modeling, multimodal representation learning

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

TransFlow uses block-wise VAEs and bidirectional Flow Matching for cross-modality translation between scRNA-seq and scATAC-seq. GenePT and Gena-LM are used to group genes and peaks of ATAC-seq into blocks. Instead of starting from random noise, the Flow Matching model starts from the source modality representation and learns to transport it to the target modality. TransFlow is evaluated on BMMC, PBMC10k, and MiniAtlas. The method achieves good performance on distribution-level metrics, while improvements in feature-level prediction are less consistent. The authors also conduct ablation studies and held-out-cell-type experiments.

Critical strengths
  1. Initializing Flow Matching from the source modality is a reasonable design for paired multi-omics data. The ablation study shows clear improvements over noise initialization.
  2. The authors evaluate the method on three datasets with several baselines. TransFlow achieves good distribution matching performance in both translation directions. The held-out-cell-type experiments also provide some evidence of generalization.
Critical weaknesses
  1. Flow Matching has already been applied to single-cell multi-omics generation and cross-modality translation, such as in CFGen and GENOT. The main contribution is source initialization, while the bidirectional architecture is largely based on existing methods. The block-wise representation also provides only marginal improvements.
  2. The evaluation mainly uses standard datasets, and comparisons with existing Flow Matching approaches are not sufficiently controlled. Improvements are mostly reflected in distribution-level metrics, while feature-level prediction is less consistent. More challenging datasets and clearer examples of generated profiles would strengthen the results.
Questions for authors
  1. Can the authors provide a more controlled comparison with existing Flow Matching approaches, using the same latent representations and comparable architectures?
  2. Can the authors evaluate TransFlow on more challenging datasets, such as cross-tissue or cross-dataset translation? The current experiments mainly use standard benchmarks.
  3. Why does TransFlow achieve better W2 and MMD scores but not consistently better feature-level prediction (such as Row Cor. in Table 1)?
  4. Since the VAEs of RNA-seq and ATAC-seq are trained independently, what justifies linear interpolation between their latent representations?
  5. Why the results for the full TransFlow model differ between Table 1 and Table 3 on BMMC?
Reviewer_6YTtOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TransFlow performs bidirectional RNA–ATAC translation by independently encoding the two modalities with block-wise VAEs and learning a continuous velocity field between paired source and target latents. Unlike conventional noise-to-data flow matching, inference begins from the observed source latent. A shared BiFM-style network handles both directions, conditioned on cell type and batch metadata.

The paper reports good distribution-level W2 and MMD results, but feature-level fidelity is mixed. In my assessment, the principal benchmark is confounded by unequal access to ground-truth metadata, while the claimed methodological novelty over existing source-to-target flow and transport formulations is limited.

Critical strengths
  • The noise-initialized and direct-regression controls attempt to isolate source initialization from source availability.
  • The evaluation includes distributional, paired feature-level, and label-transfer metrics rather than relying only on UMAPs.
  • The BMMC held-out-cell-type experiment is useful because it removes explicit cell-type conditioning.
Critical weaknesses

1. The headline benchmark gives TransFlow privileged ground-truth metadata

The benchmarking is not fair. TransFlow receives the true cell type and batch at inference. The listed configurations do not establish that most baselines receive the same information. The primary distributional metrics are then computed separately within the same true cell types. A model explicitly conditioned on those labels has a direct advantage in producing the correct class-conditional distribution. This affects W2 and MMD, not only label-transfer accuracy.

The results themselves show signs of label-driven denoising:

  • On BMMC ATAC→RNA, generated TransFlow profiles have classification accuracy 0.9248, exceeding the real-test reference accuracy of 0.8713.
  • On PBMC10k ATAC→RNA, generated accuracy is 0.9833 versus 0.9370 for real data.
  • The paper treats a negative signed accuracy gap as better, although outperforming real data can indicate removal of within-class variation or leakage from supplied labels rather than greater biological fidelity.

2. Distributional performance does not translate consistently into paired-cell fidelity

Cross-modality translation is paired: the prediction should recover the target measurement of the same cell. Yet TransFlow’s strongest results are population-level metrics, while its paired feature results are frequently inferior such as ATAC→RNA row correlation, NA→ATAC AUPRC, and RNA→ATAC AUPRC.

3. The source-initialized latent path is not theoretically well justified

The RNA and ATAC VAEs are pretrained independently. Their latent coordinates are therefore unidentifiable up to rotations, permutations, scaling, and other transformations. Consequently, the straight line

zt=(1−t)zRNA+tzATACz_t=(1-t)z_{\mathrm{RNA}}+tz_{\mathrm{ATAC}}

does not have an invariant biological or geometric meaning merely because the underlying measurements come from the same cell.

There is also a specification gap: the paper defines RNA and ATAC representations with potentially different token counts, LrL_r and LaL_a, but directly interpolates the resulting tensors. Equation 2 requires identical shapes and a defined correspondence between RNA and ATAC token positions. It is not explained whether Lr=LaL_r=L_a, how this is guaranteed, or what it means to align an arbitrarily ordered gene block with an arbitrarily ordered peak block.

4. Architectural novelty is limited

The method largely combines:

  • independently pretrained modality VAEs;
  • block construction inherited from SAVE;
  • standard linear-interpolation flow matching;
  • the shared interval-conditioned network and consistency objective from BiFM;
  • standard metadata conditioning through adaLN.

BiFM already introduces shared bidirectional interval-conditioned velocity learning and consistency supervision. GENOT already uses flow matching for transport between heterogeneous modalities, including ATAC-to-RNA translation. More generally, dynamic conditional optimal transport formulates flow-based transport between arbitrary source and target distributions, and the shared bidirectional machinery comes directly from BiFM.

Using the observed paired latent instead of Gaussian noise is sensible, but it appears to be a straightforward choice of base distribution and coupling rather than a substantial new flow-matching principle.

5. The decisive ablations are not reliable enough

Table 3 uses unexplained values that differ materially from the principal BMMC evaluation:

  • Main-table TransFlow ATAC→RNA: W2 11.2189 and MMD² 0.0095.
  • Ablation-table TransFlow ATAC→RNA: W2 9.170 and MMD² 1.026.

The paper labels both with the same metric names but does not explain the changed evaluation space or protocol. This undermines the source-initialization and continuous-flow conclusions.

6. Data splits and uncertainty remain underspecified

The manuscript says “multiple independent runs” without reporting their number or which randomness is varied. Table 3 and the held-out experiment provide no uncertainty.

The MiniAtlas description is particularly ambiguous: a random 20% subset is said to serve as the cohort for both training and subsequent evaluation, but no internal train/test split is described. This must be clarified to exclude evaluation on training cells.

Questions for authors
  1. Can you report the complete benchmark with no cell-type or batch conditioning for TransFlow?
  2. Which baselines received exactly the same cell-type and batch metadata?
  3. Are the VAE encoders and target decoders also cell-type-conditioned, as Appendix C appears to indicate?
  4. Why should linear interpolation between two independently trained and unaligned latent spaces be meaningful?
  5. Are LrL_r and LaL_a equal? How are RNA and ATAC token positions placed into correspondence?
  6. Why do Table 1 and Table 3 report substantially different TransFlow W2/MMD values?
  7. How many independent runs were used, and which components were retrained?
  8. Which cell types were held out, and were multiple held-out selections evaluated?
  9. What is the exact MiniAtlas train/test division?
  10. Does TransFlow preserve within-cell-type variation rather than generating label-specific prototypes?
  11. Do conclusions hold if ∣ΔAcc∣|\Delta\mathrm{Acc}|, rather than signed ΔAcc\Delta\mathrm{Acc}, is used?
Reviewer_sSZ6OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes TransFlow for paired RNA↔ATAC translation. Each modality is first encoded by a frozen, block-wise VAE. Features are grouped into blocks using GenePT gene embeddings (RNA) and Gena-LM sequence embeddings (ATAC). A single interval-conditioned velocity network (BiFM, Dai et al. 2026) is then trained on the straight paired path between the two latent. The network is conditioned on cc (cell type and batch) through adaLN. Classifier-free guidance is applied at inference (App. C.3). Translation is a deterministic Euler integration that starts from the encoded source cell (Alg. 2).

The claimed contributions are:

A formulation of paired translation as transport that starts from the source latent, as opposed to noise-initialized conditional generation. Block-wise representations combined with a shared bidirectional velocity field. Evaluation on BMMC, PBMC10k and MiniAtlas, covering distributional, feature-level and label-transfer metrics, plus a held-out-cell-type split and ablations.

TransFlow reports the lowest class-macro W2W_2 and RBF-MMD² in the main comparison. Its feature-level results are mixed.

Critical strengths

A simple and sensible design for paired data. Starting the flow at the source cell is natural when pairs are observed.

Broad evaluation design. The paper covers 3 datasets with distributional, feature-level and label-transfer metrics, plus a PCA-dimension sensitivity analysis (App. H). .

Detailed written documentation. The preprocessing, metric definitions (D.7) and baseline settings (D.5) are documented. Key details are still missing, however, and the linked code is unavailable (W8).

Critical weaknesses
  • pθ(xATAC∣xRNA,c)p_\theta(x^{\mathrm{ATAC}}\mid x^{\mathrm{RNA}},c) (Eq. 1) is a deterministic map. The ODE preserves the target marginal, not the observed pairing. This fits the results: strong distributional metrics, weak per-cell metrics (Table 1a). Either sample and use proper scoring rules, or present TransFlow as a point predictor.

  • The comparison appears potentially unmatched in conditioning: TransFlow explicitly receives cell-type (and BMMC batch) labels, while equivalent auxiliary inputs are not documented for the other translation baselines. Please report a no-cell-type TransFlow comparison and clarify conditioning for every baseline.

  • Inconsistent numbers: some MMD² values apear to exceed the bound implied by Eq. 13 (Tables 1c, 2, 3). Table 3's TransFlow row does not match Table 1a. Several stds are zero, and the number of runs is unstated.

  • Baselines and ablations need stronger validation. scMoGNN performs close to chance on the label-transfer metrics, which raises concern about whether the baseline is functioning as intended. The Direct Latent MLP also performs very poorly. Since a one-step TransFlow is effectively a residual latent prediction, please report a one-step TransFlow ablation. GENOT is run in pure GW mode even though paired cells are available. A comparison using the observed pairing would be more informative.

  • Underspecified experiments: Eq. 2 needs Lr=LaL_r=L_a, which is never established; Fig. 2a suggests otherwise.

  • Full pairing is required. The paper is motivated by unimodal cells, yet every training cell must be paired. The common few-paired/many-unpaired case is ignored.

  • Code unavailable and thus paper not reproducible (see also the other comments). The anonymous repository (App. A) has expired.

Questions for authors

Does TransFlow model a distribution? Please report per-cell sample diversity and a proper scoring rule (e.g. energy score).

Can you reconcile the MMD² values with Eq. 13, and Table 3 with Table 1a?

How many runs are there, and what varies between them?

How does the Noise-Initialized FM receive the source? Why does the Direct MLP perform so poorly?

Which cell types are held out, and how does the probe classify unseen classes?

Can GENOT use the observed pairing, and can scMoGNN be fixed?

What happens with λcon=0\lambda_{\mathrm{con}}=0, and without the t′t' input?

What are LrL_r and LaL_a, and how are they made equal?

How does performance change when only part of the training cells are paired?

How does TransFlow compare with aligned Schrödinger bridges (Somnath et al. 2023, Aligned Diffusion Schrödinger Bridges, https://arxiv.org/abs/2302.11419)?

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

Related papers

Discussion (0)

Sign in to comment.