acceptodds
Under review as a conference paper at ICLR 2027

Dual-Track Flow Matching for Generative RNA Base Pair Prediction

Abstract

RNA molecules can adopt multiple base-pairing states, yet most learning-based predictors return a single structure per sequence. We introduce **DTFlow**, a dual-track flow framework that unifies prediction and generation of sparse binary base-pair matrices (BPMs), including non-canonical interactions. A pretrained sequence track and a pair track exchange information through attention at nucleotide-pair resolution. Two time step variables condition this exchange and, together with the input states, gate prediction, conditional BPM generation or auxiliary masked sequence reconstruction within the same architecture. Prediction and generation share a BPM endpoint readout and objective, enabling predictive training to initialise generation from limited structural supervision. We evaluate this framework with **RNA-MultiBP**, which preserves multiple observed BPMs per sequence and excludes target-related sequence groups from training and validation through sequence-, structure- and family-informed clustering. On 56 CASP and RNA-Puzzles sequences, we assess reference coverage and candidate fidelity by comparing predicted and observed base pairs. DTFlow's single-output predictor improves reference coverage over the RiNALMo baseline by 11.1 percentage points, while approaching RNAfold and ProbKnot. With 16 candidates, its generative model improves coverage over retrained RNADiffFold by 9.6 percentage points at the same candidate count. On the 26 multi-reference sequences, generating 16 candidates improves coverage over the single predictive output by 3.1 percentage points. Conditioning trRosettaRNA2 on DTFlow samples brings predictions of a riboswitch closer to deposited conformers than conditioning on the deposited BPMs themselves. These results demonstrate competitive prediction and improved candidate-set coverage within a unified framework under limited structural supervision and controlled sequence separation. Code is available at [https://anonymous.4open.science/r/DTFlow-3C0D/](https://anonymous.4open.science/r/DTFlow-3C0D/).

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
RNA base pair, Conditional generative modeling, Flow matching, Multi-conformation, Dual-track

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_49djOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The authors introduce DTFlow for generation and prediction of RNA base-pair matrices (BPMs). Recognizing the importance of modeling RNA's multi-state conformations, the DTFlow architecture prioritizes BPM representation learning through its dual RNA sequence-pair representation channels. Empirical results suggest that DTFlow is modestly competitive for RNA BPM prediction across different biological contexts.

Critical strengths
  • The authors identify the importance of modeling RNA BPMs for multi-state RNA conformation prediction.
  • DTFlow borrows useful ideas from the protein structure prediction literature for robust RNA modeling, including pair-bias attention.
  • DTFlow is evaluated comprehensively for RNA BPM prediction.
Critical weaknesses
  • Although the use of large language models (LLMs) is certainly not prohibited in preparing ICLR manuscripts, my reading of this paper suggests that the authors heavily relied on LLMs for code development, experiment design, and manuscript writing, resulting in a very dense 9-page block of main text that I found difficult to parse. I would have liked to see the authors take the time not to use verbatim LLM text in their submitted manuscript, as I believe this would have greatly improved the readability of the manuscript. As two examples of (what I believe to be) vague LLM writing in the main text, Lines 86-88 ("Evaluation measures reference coverage and candidate fidelity at explicit candidate counts: how closely the candidate set reaches each reference, and each candidate matches an observed BPM") and Lines 278-279 ("Split construction excludes target-related sequence groups using sequence, structural and fragment-reuse links") are practically unreadable to me.
  • Given that the authors have (presumably) relied heavily on LLMs to submit this manuscript, for me, it calls into question what I'm supposed to take away from this work, given that DTFlow's results across different benchmarking contexts (e.g., Table 1 - "single", "inter", "CASP15", and "CASP16" columns) are mixed compared to "Thermodynamic" baselines and that this manuscript is difficult to read in the first place.
  • Lastly, given that DTFlow seems to offer modest performance advantages in a small number of RNA-specific contexts without being described clearly, I find it difficult to draw any notable insights from reading this work, such as compelling representation learning or generative modeling ideas to consider in future work on AI for biology.
Questions for authors
  • Why do the authors use 1x1 convolutions instead of making DTFlow entirely based on attention? If the authors had used attention only, then they might have been able to leverage optimized CUDA kernels to accelerate training and inference (e.g., cuEquivariance). The rationale for choosing 1x1 convolutions was not clear to me in this light.
  • Would the authors please consider replacing the term "trails" with another verb, since its usage throughout the manuscript confused me quite a bit (e.g., does it carry a positive or negative connotation in all contexts here)?
Reviewer_vu4mOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the task of predicting an RNA base-pair matrix (BPM) from sequence alone. BPM targets record annotated pairs under any Leontis–Westhof geometry, including non-canonical ones, and unresolved cells are masked out rather than treated as non-pairs. A RiNALMo encoder is coupled to a full-resolution pair track through pair-biased attention; two time variables select between BPM prediction, conditional BPM generation, and masked sequence reconstruction, with the predictive weights initialising the generative stage and generation done by Euler steps from a uniform source. The paper also contributes RNA-MultiBP, a PDB-derived benchmark that keeps multiple observed BPMs per sequence and scores candidate sets by coverage and fidelity on 56 test RNAs. DTFlow-P reaches 0.634 All coverage, against 0.641 for RNAfold and ProbKnot and 0.523 for a RiNALMo structure head; DTFlow-G at K=16 reaches 0.651, against 0.555 for their retrained RNADiffFold. A riboswitch case study reports that conditioning trRosettaRNA2 on generated BPMs yields candidates closer to each of three deposited conformers than conditioning on the deposited BPMs.

Critical strengths
  1. The multi-reference evaluation protocol is a real contribution. RNA-MultiBP keeps every distinct observed BPM per sequence instead of collapsing them to one reference. Uncalled cells are treated as unknown rather than as non-pairs (A.1), and scoring is set-based (coverage, fidelity) rather than a single F1. This is what survives replacing the model, since current RNA benchmarks score one structure against one reference; Table 9 shows it works in practice.

  2. The paper proposes a testable mechanism for why generative base-pair prediction matters. Conditioning trRosettaRNA2 on four generated BPMs reaches TM 0.85/0.80/0.62 against three deposited conformers, versus 0.57/0.57/0.39 for the predictor's own BPM. The paper concludes that "a single input map, even a correct one, constrains the predictor to one preferred static fold" (Section 4.5).

Critical weaknesses
  1. The generative advantage is not attributable to generation. Coverage takes the maximum over candidates, so it cannot fall as K grows; DTFlow-G at K=16 is nevertheless compared with DTFlow-P at K=1, with no matched-budget control. At equal K on the authors' own common cohort the released RNADiffFold reaches higher All coverage (0.660 vs 0.640, E.6), so the abstract's "+9.6 points at the same candidate count" holds only against the retrained baseline (0.555). Section D.2 also states that the validation selection for the reported (T,K) = (2,16) is unverified.

  2. The methodological core is not shown to contribute. Under the paper's own endpoint parameterisation, T=1 sampling is a single forward pass from a random source, and the results agree that more steps do not help (0.645/0.624 at T=1 versus 0.651/0.591 at T=2, K=16). The gain over the RiNALMo head is confounded because DTFlow-P alone receives 20,000 RNAfold-labelled mutants. The dual-time weighting is never ablated — the Limitations section concedes this — so the paper describes the dual-track design but shows no experiment in which either component contributes.

Questions for authors
  1. Table 1 reports coverage only, and coverage takes a maximum over candidates, so it cannot fall as K grows; fidelity appears only in Table 5. Please add fidelity to Table 1 and mark K and T per row, so the main table can be read on its own.

  2. Table 1 and Table 3 use different DiT checkpoints for the same configuration (0.150 vs 0.232, D.3). Please use one checkpoint in both tables, or label them as two separate models. The retrained RNADiffFold also keeps residual UFold normalisation tensors after reset (D.3); please retrain with those removed, or use the released checkpoint already reported in E.6.

  3. Two results that bear on the headline appear only in the appendix: ViennaRNA centroid reaches All coverage 0.656 with one candidate, above DTFlow-G's 0.651 with 16 candidates (E.5), and the released RNADiffFold reaches 0.660 against 0.640 at equal K on the common cohort (E.6). Why are these rows not in Table 1? Please add them to the main comparison.

  4. Three controls are needed to support the central claims. (i) Matched budget: sample K=16 from the stage-1 predictor with a random pair source at (t_seq, t_pair) = (1, 0). Does it reach DTFlow-G's coverage and fidelity? Also report the single-output one-to-one match count on the 85 Multi groups under the Table 9 criterion. (ii) Matched supervision: DTFlow-P alone is trained on 20,000 RNAfold-labelled mutants; please train the RiNALMo structure head with the same augmentation, or report DTFlow-P without it. (iii) Training-level ablations: the feedback result only gates at inference, so please retrain without the pair-to-attention feedback and without the dual-time weighting. The tertiary comparison in Section 4.5 also uses unequal candidate budgets (D.4); please match them.

  5. Selection and training details are missing. D.2 states that the validation selection for the reported (T,K) = (2,16) is unverified — was this setting chosen on the test sweep? The learning rate, batch size, update count, w_seq, task mixture probabilities and the sampling distribution of (t_seq, t_pair) are not reported, although C.3 says they must accompany each training configuration. If t = 1 carries no probability mass, the clean-BPM conditioning input of Both is a measure-zero input. Please release the manifest and score exports named in D.1, and report how much the reference groups of one sequence actually differ — at τ = 0.3, Table 9 already matches 83 of 85 groups.

  6. Symbols are inconsistent: Table 2 uses u and s for the BPM and sequence corruption times, while the method uses t_pair and t_seq, and s already denotes the native sequence in Eq. (5). Please unify them. Please also confirm (i) the identical RNAfold and ProbKnot values in Table 5 (0.641 coverage, 0.647 fidelity on All, though they differ on Multi), (ii) the DiT-BPM values in Table 1 (0.150) and Table 3 (0.232), and (iii) the abstract's 3.1 percentage points against the 3.0 implied by Table 1 (0.653 vs 0.623).

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

This paper proposes Dual-Track Flow Matching (DTFlow), a generative approach to RNA base-pair prediction that jointly models sequence representations and base-pair maps. It combines sequence-conditioned prediction, flow matching over pair maps, and feedback between the two tracks. The authors also introduce RNA-MultiBP, a benchmark that preserves multiple observed structures and no-call regions, and describe split controls intended to reduce sequence and structural leakage. On a test set of 56 RNA sequences, DTFlow-G with T=2,K=16T=2, K=16 reports higher all-set coverage than the retrained RNADiffFold baseline, while DTFlow-P is competitive with thermodynamic methods. The paper includes analyses of coverage, fidelity, inference feedback, and limitations in the evaluation and implementation.

Critical strengths
  1. The benchmark design takes annotation and leakage issues seriously. RNA-MultiBP preserves observed base-pair annotations and no-call masks, groups identical sequences, and documents sequence-, structure-, and fragment-aware split controls. These choices make the benchmark more useful for evaluating multi-state predictions than a setup that collapses observations into a single reference structure.

  2. The evaluation examines both coverage and fidelity and includes meaningful caveats. The paper distinguishes matching observed structures from recovering every biological state, discusses cases where thermodynamic methods perform better, and reports limitations of the current evaluation. This makes the empirical picture more informative than a single aggregate score.

Critical weaknesses
  1. The selection provenance and uncertainty around the headline comparison remain unclear. The main result uses DTFlow-G with T=2,K=16T=2, K=16, but Appendix D.2 says that a completed validation-selection record for this default “has not yet been verified,” and cautions against inferring it from the test-set optimum. This does not establish that test results were used for selection, but it leaves the separation between configuration choice and final evaluation difficult to verify. The test set contains only 56 sequences, including 26 multi-reference sequences, and the main comparisons are reported as point estimates without confidence intervals (Appendix B.5). This matters especially for relatively small differences, such as DTFlow-G’s all-set coverage of 0.651 versus RNAfold’s 0.641. Even if the configuration was selected using validation data, paired sequence-level and, where appropriate, family-grouped uncertainty estimates would help establish how robust these differences are.

  2. The experiments do not cleanly isolate the proposed architectural contribution. The two-stage configurations in Table 2 combine predictive initialization with additional predictive-stage training and 20,000 RNAfold-labelled sequence mutants. The reported gains therefore support the overall training recipe, but do not isolate the contribution of predictive initialization or cross-track coupling. The inference-feedback ablation is also limited: it uses a decoding threshold of 0.15, whereas the main grid uses 0.3; its all-set coverage change is only +0.0057, with a reported 95% bootstrap interval that includes zero. Appendix E.4 further reports at most two additional one-to-one matches, with intervals that include zero. Finally, Appendix D.3 notes residual supervised normalization tensors in the RNADiffFold initialization after reset, whose effect cannot be determined without a clean reinitialization control. These points make it difficult to attribute the observed gains specifically to DTFlow’s proposed components.

Questions for authors

Q1: Can the authors provide the validation-selection record for T=2,K=16T=2, K=16, including which choices were fixed before test evaluation? Could the main comparisons also be reported with paired sequence-level uncertainty estimates and family-grouped estimates where relevant, particularly for the small DTFlow-G versus RNAfold difference?

Q2: Can the authors separate the effect of predictive initialization from the additional predictive training and RNAfold-labelled mutants in the Table 2 two-stage results? Could the feedback ablation be reported at the main decoding threshold, and could the RNADiffFold comparison be repeated or otherwise controlled with a clean reinitialization? These answers would materially affect how much of the gain I attribute to the proposed dual-track coupling rather than to the broader training recipe or baseline initialization.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.