- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper introduces sCM-RNA, a one-step RNA inverse-folding framework that combines structure-conditioned flow matching, continuous-time consistency distillation, and structure-guided preference alignment. A flow-matching model is first distilled into a one-step consistency model, after which RhoFold-based structural preferences and entropy regularization are used to improve structural fidelity while preserving sequence diversity. The method is evaluated on the standard RiboDiffusion split and a recent CAMEO OOD set, with comparisons to GVP-GNN, gRNAde, RiboDiffusion, and the flow-matching teacher. The paper reports improvements in TM-score, recovery, and diversity, together with substantially lower inference latency, and further evaluates robustness using Chai-lab, multiple random seeds, pass@16, and additional ablations on the preference and diversity objectives.
Critical strengths
-
The paper tackles an important practical bottleneck in generative RNA design by combining one-step consistency generation with structure-guided preference alignment. This gives a favorable quality–efficiency trade-off: the model improves structural fidelity and diversity while reducing generation latency by more than an order of magnitude relative to multi-step diffusion/flow baselines.
-
The empirical validation is relatively comprehensive. Results are reported over six seeds, on both the standard in-domain split and a recent CAMEO OOD set, and the structural gains are further checked with Chai-lab and fixed-budget pass@16 evaluation. This makes the reported improvements more convincing than relying on a single folding oracle or a single in-domain benchmark.
Critical weaknesses
-
The Stage-II distillation result is difficult to interpret. Before preference alignment, the one-step student sCM (Table 5: Recovery = 0.5382, TM-score = 0.5956) already outperforms the multi-step flow-matching teacher (Table 1: Recovery = 0.5113, TM-score = 0.5875), while diversity drops from 0.3040 to 0.2818. Since Stage II uses a simplified tangent-residual formulation rather than the original sCM recipe, and the paper neither compares against standard sCM nor varies the teacher sampling steps, it is unclear whether this gain comes from the modified distillation objective, finite-step solver error in the teacher, or simply a more concentrated output distribution.
-
The justification for the sCM-DPO objective is also not fully convincing. While inference only queries , sCM-DPO samples without the consistency loss, and the endpoint reconstruction error scales with (vanishing at ). The reported 76% ranking agreement in Appendix G.1 is measured only on filtered high-margin extreme pairs at , which does not validate the surrogate across the training time distribution. Moreover, the per-position probability in the token-entropy regularizer in Eq. 8 is not mathematically defined for the continuous non-autoregressive predictor.
-
It seems like numerical inconsistencies across tables and missing aligned baselines. Purely sequence-level metrics (Recovery, MFE, Diversity) are independent of the 3D folding oracle, yet they vary for the same models across Table 1 (RhoFold), Tables 2–3 (Chai-lab), and Table 5. In addition, baseline TM-scores (gRNAde, GVP-GNN) in Table 1 (0.5032 / 0.5601) differ substantially from Appendix Table 9 (0.2932 / 0.4101), and the FLOPs reduction ratio in Figure 4(b) conflicts with Table 9. Finally, the paper omits comparisons with preference-aligned baselines such as RiboPO or applying DPO directly to the flow-matching teacher.
Questions for authors
-
Could the authors directly compare the proposed Stage-II distillation with the original sCM training formulation under the same architecture and training budget, and test whether the one-step student still outperforms the flow-matching teacher when the teacher is evaluated with more ODE solver steps?
-
The base sCM improves TM-score and recovery but reduces sequence diversity relative to the flow-matching teacher, whereas Stage-III alignment subsequently reduces recovery. Could the authors clarify whether the apparent pre-alignment quality gain is mainly due to a more concentrated output distribution?
-
How well does the endpoint reconstruction error preserve the RhoFold preference ordering across the full time distribution used during sCM-DPO training? Additionally, how is in the entropy regularizer in Eq. 8 mathematically computed from the continuous predictor ?
-
Could the authors clarify why purely sequence-level metrics (Recovery, MFE, Diversity) differ across Tables 1, 2, 3, and 5 for the same models, and why baseline TM-scores and FLOPs ratios differ between the main text and Appendix Table 9?
-
Tables 3 and 6 evaluate RhoFold-aligned models with Chai-lab, while Appendix G.2 only evaluates Chai-lab-aligned models with Chai-lab. Could the authors complete the cross-oracle evaluation by testing the Chai-lab-aligned model with RhoFold, and compare against applying DPO directly to the flow-matching teacher?