acceptodds
Under review as a conference paper at ICLR 2027

La-Ribo: RNA Co-Design via Geometry–Latent Flow Matching

Abstract

RNA function arises from the coupling of nucleotide sequence and three-dimensional structure, motivating their joint design. Coordinating global folding with nucleotide-level detail remains challenging under limited structural supervision. We introduce La-Ribo, a generative framework for RNA sequence-structure co-design via geometry-latent flow matching. La-Ribo retains a sparse phosphate–sugar–base scaffold and encodes nucleotide identity and local conformation in residue-wise latents. A shared flow network generates both jointly, and an RNA-specific decoder then reconstructs all heavy atoms. To expand supervision, we construct a quality-controlled corpus of 168,561 RNA structures, integrating experimental data with predictions from three folding models, including 10,631 MSA-supported structures generated in this work. La-Ribo improves designability and codesignability over the evaluated baselines across sampling budgets and two refolding models, and the same prior supports scaffold-conditioned inverse folding without additional training. Our code is available at https://anonymous.4open.science/r/la-ribo-7DDA.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Generative Model, Biomolecule, RNA Design

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_EwQ6OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The authors proposed La-Ribo, a generative pipeline for RNA sequence (co)-design, via geometry–latent flow matching. Besides the framework itself, this submission also includes a novel corpus of 168,561 RNA structures across experimental results and in silico generated data from OpenFold3, NuFold and trRosettaRNA2. In general, this is a comprehensive study with open-source code and most of the statements documented in this submission are supported by the quantitative results of various experiments.

Critical strengths
  1. A new dataset towards precision RNA generation (design) was included. Specifically, the authors managed to integrate both experimental and in silico structural data from different sources.
  2. A comprehensive appendix was well documented. Besides the promising dataset mentioned above, several general concerns (e.g., computational cost, details of the training strategy) have been addressed quantitatively.
Critical weaknesses
  1. It is still UNCLEAR for me that why the three atoms (i.e., P, C4' and glycosidic N) are picked as sparse geometric anchors. Some supporting information needs to be added afterwards, to specify the reasons behind it. Compared with the similar strategy (i.e., C, Cα and N), widely applied in the field of protein design, I assume the authors are trying to identify the backbone atom(s) of each RNA base, which is pretty interesting and meaningful. However, something more efforts have to be made. Specifically, besides the finalized strategy and those options included in Table 11, have the authors ever tried any other amendments?

  2. I notice that the authors chose to generated another batch of in silico data via trRosettaRNA2, based on the sequences from bpRNA-1m(90) and RNAStralign. I'm wondering if the authors ever applied other existing folding pipeline(s) to generate the corresponding 3D structures from the original sequences. If the answer is yes, I would be so happy if the authors could post the detailed comparisons afterwards. If the answer is no, several representative pipelines (e.g., AlphaFold3, Chai-1, Protenix) are highly recommended to be used to generated the 3D structural data. Additionally, a follow-up concern regarding those entries from NuFold: which species have been included in the final dataset?

Questions for authors

Please refer back to "Critical Weaknesses".

4 comments
AuthorsOpenReview ↗

Response to Reviewer EwQ6 (1/2)

We thank the reviewer for the positive assessment and the constructive questions. In brief:

  • Q1. The anchors follow the chemistry of a nucleotide and the frame the atom builder needs; retrained variants with other anchor sets or without a learned rotation are all weaker than P/C4′/N with a learned rotation.
  • Q2. For the bpRNA-1m(90)/RNAStralign subset we have so far used trRosettaRNA2 only; we plan to release structures from other predictors such as Protenix.
  • Q3. The NuFold subset covers 47 species, listed below.

Q1. Why P, C4′ and glycosidic N, and have other options been tried?

Design rationale.

  • Each anchor locates one chemically distinct part of a nucleotide. P is the phosphate backbone (the P–C4′ trace gives the standard pseudotorsional description of the backbone), C4′ is the sugar and the translation origin of our decoder, and the glycosidic N (N9 for A/G, N1 for C/U) is the base attachment point. The choice follows from this chemistry, not from analogy with the protein C/Cα/N triplet.
  • The anchor frame and the atom-building frame differ, so we learn the transformation. The atom builder uses the ribose frame Ric=F(xiC4′,xiO4′,xiC3′)R^{c}_i = F(x^{\mathrm{C4'}}_i, x^{\mathrm{O4'}}_i, x^{\mathrm{C3'}}_i), whereas the P/C4′/N anchors define a different frame Rip=F(xiC4′,xiνi,xiP)R^{p}_i = F(x^{\mathrm{C4'}}_i, x^{\nu_i}_i, x^{\mathrm{P}}_i). The rotation between them varies across residues (by 9.3∘9.3^{\circ} on average from its mean), so the decoder predicts a residue-dependent QiQ_i with Ridec=RipQiR^{\mathrm{dec}}_i = R^{p}_i Q_i (Sec. 3.2).
  • The ribose atoms O4′ and C3′ are not used as anchors because they describe local ribose geometry but do not specify the positions of P and the glycosidic N, which carry the backbone trace and the base attachment.

What we tried.

For this rebuttal we retrained the autoencoder (encoder and decoder) and the flow model for the following variants, with all other settings identical to Table 1, using the lighter Base architecture and RhoFold scoring:

  • Fixed Q. Same P/C4′/N anchors, but QQ is one constant rotation: the per-residue targets (Rip)⊤Ric(R^{p}_i)^{\top}R^{c}_i are averaged over all training residues and projected onto SO(3)SO(3).
  • C4′/O4′/C3′ (identity Q). These three atoms already span the ribose frame, so no transformation is needed. It contains no phosphate or base-attachment anchor.
  • C4′ only, P/C4′, P/C4′/C3′ (learned Q). Fewer or different anchors, with QQ learned as in our model.

RhoFold, 200 steps

AnchorsQDesign. [email protected]↑Design. [email protected]↑Codesign. [email protected]↑Codesign. [email protected]↑
P/C4′/N (original model)learned38.22±1.2753.00±0.6732.56±2.2556.72±1.78
P/C4′/Nfixed32.28±2.4450.44±1.8027.17±0.7651.72±1.07
C4′/O4′/C3′identity27.94±0.6741.00±0.7329.94±1.7150.83±2.84
C4′learned24.28±0.7540.22±2.4125.17±1.7347.11±0.84
P/C4′learned22.44±0.8435.83±2.9625.11±0.3542.67±1.80
P/C4′/C3′learned31.83±3.1947.11±2.2530.94±1.7553.72±0.96

Percentages; mean ± SD over three 600-structure sampling campaigns. The first row is the original model of Table 1; the other rows are retrained variants.

Results.

In this setting (Base architecture, 200 steps, RhoFold), P/C4′/N with a learned QQ gives the highest value in every column.

  • Learning QQ matters. Freezing QQ lowers designability from 38.22 to 32.28 and codesignability from 32.56 to 27.17, and the decoder becomes less accurate: with a constant QQ the frame error is 14.2∘14.2^{\circ} and the non-scaffold RMSD 0.330 Å, against 1.78∘1.78^{\circ} and 0.087 Å for the learned QQ (Table 10).
  • Ribose-only anchors give lower designability. The identity-QQ variant reaches 27.94 designability ([email protected] 41.00) against 38.22 (53.00), consistent with the loss of the information carried by P and N.
  • Other anchor sets are also weaker. C4′ only, P/C4′ and P/C4′/C3′ reach 22.44 to 31.83 designability against 38.22. Replacing the base-attachment anchor N by the ribose atom C3′ costs several points of designability.
AuthorsOpenReview ↗

Response to Reviewer EwQ6 (2/2)

Q2. Other folding pipelines for the bpRNA-1m(90)/RNAStralign structures.

For this subset we have so far completed structure prediction only with trRosettaRNA2. We plan to generate and release synthetic structures for this subset with other competitive predictors, such as Protenix. Given the data volume, we cannot complete this comparison within the rebuttal period, but we will release the corresponding datasets to the RNA community and provide a detailed comparison at that time.

Q3. Which species are in the NuFold subset?

The full NuFold database covers 48 species, and the final dataset retains 47 of them, obtained by mapping the source taxonomy identifiers to species level through NCBI Taxonomy. The species are: Danio rerio, Homo sapiens, Rattus norvegicus, Enterococcus faecium, Mus musculus, Zea mays, Glycine max, Oryza sativa, Schistosoma mansoni, Arabidopsis thaliana, Trichuris trichiura, Drosophila melanogaster, Escherichia coli, Caenorhabditis elegans, Strongyloides stercoralis, Brugia malayi, Shigella dysenteriae, Sporothrix schenckii, Salmonella enterica, Paracoccidioides lutzii, Pseudomonas aeruginosa, Histoplasma capsulatum, Candida albicans, Saccharomyces cerevisiae, Onchocerca volvulus, Madurella mycetomatis, Wuchereria bancrofti, Nocardia brasiliensis, Neisseria gonorrhoeae, Mycobacterium tuberculosis, Dictyostelium discoideum, Haemophilus influenzae, Mycobacterium ulcerans, Dracunculus medinensis, Mycobacterium leprae, Plasmodium falciparum, Staphylococcus aureus, Campylobacter jejuni, Helicobacter pylori, Methanocaldococcus jannaschii, Trypanosoma cruzi, Leishmania infantum, Fonsecaea pedrosoi, Trypanosoma brucei, Cladophialophora carrionii, Klebsiella pneumoniae and Streptococcus pneumoniae.

Reviewer_EwQ6OpenReview ↗

Re: Authors' Responses

Dear Authors,

I would love to say thank you for the detailed responses towards most of my concerns.

I hereby confirm you all that I WILL read each and every word of yours while considering my final recommendation.

Please feel free to post some more supporting information (if applicable) during the rebuttal period.

Reviewer EwQ6

AuthorsOpenReview ↗

Thank you for your note

Dear Reviewer EwQ6,

Thank you very much for your kind note. For convenience before the discussion period closes, we also posted a brief summary on the forum ("General Response to Reviewers and AC") that collects all rebuttal results in one place, including the scaffold and rotation ablations for your Q1 and the parameter-matched Base/Tri comparisons from the other threads.

Please let us know if there is anything else we can clarify.

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

The paper introduces La Ribo - a generative model that applies the La Proteina approach to the RNA, with adjustments as needed. It also introduces new (mostly synthetic) data that is used in training and evaluation. The end result is a highly performant model in an application space where structure-sequence map is not yet maturely known.

The paper contributes

  • a representation that works for RNA by using explicit anchors.
  • a new dataset that is largely synthetic, that improves model performance, not only for La Ribo but for baseline as well, pointing at the data limit of this application space.
Critical strengths
  • a new geometric latent representation that is shown to be beneficial for generative tasks for RNA. The paper's ablation about the representation is sound.
  • a new data corpus that is mostly synthetic. It is meaningful that the synthetic data pool is generated with a different model than the evaluator. It is appropriate that authors ablate the impact of data independently of the representation and show that data helps also the baseline model.
  • they test a base model and a triMul version and show the benefit of the geometric update (although there may also be confounding of number of parameters)
  • novelty and diversity of the generated structures are also better than the compared baselines.
Critical weaknesses
  • Probably the main weakness of the work is the fact that it relies on in silico metrics generated by the structure prediction models that are known to be worse performers than they are for proteins, and even in protein generation wet lab validation is still needed.
  • Secondly, the architecture explorations seem to impact diversity only marginally, -Tri variation being very slightly exceeding existing models, -base model very slightly regressing. Comment on this would be appreciated. (I see in appendix that the OpenDDE results are more favorable instead)
Questions for authors

Please see comments on the weaknesses section

4 comments
AuthorsOpenReview ↗

Response to Reviewer ikdr

We thank the reviewer for the positive assessment and the careful comments.

1. Reliance on in-silico metrics. We agree that designability and codesignability are computational measures of sequence-structure self-consistency, and that experimental validation is the ultimate test. We are collaborating with wet-lab partners and hope to present experimentally validated results in future work; we will state this scope in the Limitations.

2. Effect of the architecture on diversity. We agree that with RhoFold the cluster-count differences are modest: at 200 steps La-Ribo-Tri has 27.67 clusters against 26.00 for RNA-FrameFlow (+1.7), and La-Ribo-Base 23.00. With OpenDDE they are larger (Tri 39.67, RNA-FrameFlow 35.00, +4.7; Base 31.33; Table 19). Three points help to read this.

  • Within the same representation, Tri has more clusters than Base. This holds in all 16 comparisons of Tables 12 and 19 (4 step counts, 2 refolders, 2 evaluation routes), by 2.0 to 12.0 clusters; for designability at 200 steps it is +4.7 clusters (+20%) with RhoFold and +8.3 (+27%) with OpenDDE. This supports a contribution of the triangle updates to the structural coverage of the structures that pass the filter.

  • Parameter count does not explain the structure-level gains. La-Ribo-Tri has more parameters than La-Ribo-Base (65.41M vs 58.82M), so we trained two variants that change only the token width of the flow network, and evaluated them with the same sampling settings as the paper's models: La-Ribo-Tri-Slim (token width 480, 58.52M, slightly fewer parameters than Base) and La-Ribo-Base-Wide (token width 544, 66.05M, slightly more parameters than Tri).

    RhoFold, 200 steps

    ModelParamsDesign. [email protected]↑Design. [email protected]↑Codesign. [email protected]↑Codesign. [email protected]↑
    La-Ribo-Base (paper)58.82M38.22±1.2753.00±0.6732.56±2.2556.72±1.78
    La-Ribo-Tri (paper)65.41M39.11±2.1258.83±1.4832.06±0.7557.89±1.71
    La-Ribo-Tri-Slim58.52M37.39±0.9255.89±2.4129.17±0.6750.67±3.00
    La-Ribo-Base-Wide66.05M32.61±0.1048.72±0.7925.67±1.0149.72±0.98

    Mean ± SD over three 600-structure sampling campaigns; paper rows from Tables 1 and 12. At matched size, Tri-Slim against Base gives designability −0.8 and TM +2.9, and Tri against Base-Wide gives +6.5 and +10.1.

  • The count depends on the refolder. #Clust is computed on the structures whose refolding passes scTM ≥ 0.45. The generated backbones are identical under both refolders, but the set that passes differs, so the gap differs between RhoFold and OpenDDE.

AuthorsOpenReview ↗

Follow-up on our response to Reviewer ikdr

Dear Reviewer ikdr,

Thank you again for your positive assessment and helpful comments. Before the discussion period closes, we wanted to check whether our response addresses your points, including the diversity comparison across RhoFold and OpenDDE, the parameter-matched variants (Tri-Slim vs. Base and Tri vs. Base-Wide), and our discussion of the in-silico evaluation scope.

Please let us know if any further clarification would be helpful.

Thank you for the responses. I appreciate the ablation for Tri ops parameter count. This confirms the intuition that geometric operations help network performance for structure generation tasks, as it was also for the La Proteina model. It is reassuring to see the trend does not depend on biomolecule modality.

I would keep my score at an accept but I hesitate to increase to a strong one, since the impact is hard to judge without wetlab validations that build trust to these validation metrics for ribosome structures. If the in silico folding methods were better established perhaps reliance on them would be a good choice but they too seem to be very early in their progress for this modality (hence big difference in the results depending on the folding model). The diversity metrics are perhaps more reliable but in those the model does not bring a step change, so I will keep my score and thank the authors for their work.

AuthorsOpenReview ↗

Thank you for your follow-up

Dear Reviewer ikdr,

Thank you for engaging with our responses and for your continued support. We agree that wet-lab validation is the key next step for building trust in in-silico metrics for RNA; this is also why we report results with two refolders, and we will make this scope explicit in the Limitations. Experimental validation with our wet-lab collaborators is underway, and we look forward to sharing those results in future work.

Compared with proteins, RNA structure modeling is still at an early stage, and we hope La-Ribo, together with the 168K-structure corpus and open-source code, can serve as a useful resource for the RNA design community.

Thank you again for your time and constructive feedback.

Reviewer_Ub79OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes La-Ribo, a geometry–latent flow matching framework for RNA sequence–structure co-design. The method combines an explicit geometric scaffold with continuous residue-wise latents encoding nucleotide identities and local atomic details. A two-stage training procedure first learns this representation through an autoencoder and then trains a joint flow matching model. The authors also construct a large RNA structural dataset and demonstrate improvements in designability and codesignability over the evaluated baselines.

Critical strengths
  1. The paper demonstrates promising empirical results, with consistent improvements in designability and codesignability across different sampling budgets and refolding models.
  2. The RNA-specific scaffold representation and structured decoder provide a reasonable solution for reconstructing all-atom RNA structures from compact representations.
  3. The construction of a curated dataset containing 168,561 RNA structures is valuable, particularly given the scarcity of experimentally resolved RNA structures.
Critical weaknesses
  1. Limited methodological novelty compared to La-Proteina. My main concern is that the proposed framework is highly similar to La-Proteina (ICLR 2026). Both approaches employ a partially latent representation that combines explicit geometric coordinates with continuous residue-wise latents, followed by a two-stage training procedure consisting of an autoencoder and joint geometry–latent flow matching. The use of modality-specific time schedules and continuous latent representations for nucleotide or amino acid identity also closely follows the existing framework. The primary differences appear to be the RNA-specific geometric scaffold and the scaffold-to-sugar frame transformation used for atom reconstruction. While these adaptations are technically reasonable, they do not seem to introduce substantial methodological innovations beyond the existing framework. I would appreciate a clearer explanation of the fundamental modeling challenges that distinguish La-Ribo from a relatively direct RNA adaptation of La-Proteina.
  2. The necessity and effectiveness of the RNA-specific representation could be better justified. A central contribution of the paper is the choice of three geometric anchors (P, C4′, and glycosidic N) together with a learned transformation to the canonical sugar frame. The authors motivate this design through RNA-specific geometric constraints and provide reconstruction ablations with different numbers of anchors. However, these experiments primarily demonstrate improvements in autoencoder reconstruction accuracy, rather than the downstream benefit of the proposed representation for generative modeling. It remains unclear to what extent the particular scaffold design and learned frame transformation are essential for the reported improvements in designability and codesignability. A more direct comparison between alternative representations under the same generative modeling setup would strengthen the justification for this design.
  3. Some architectural improvements are not sufficiently isolated. The paper introduces La-Ribo-Tri, which incorporates triangle multiplicative updates to refine pair representations. While the results show improvements in several metrics, La-Ribo-Tri also has more parameters than La-Ribo-Base (65.4M versus 58.8M). Without a parameter-matched comparison, it is difficult to determine whether the improvements arise from the proposed architectural design or increased model capacity. Additionally, the results are not uniformly better across all metrics. For instance, La-Ribo-Base achieves slightly higher strict codesignability than La-Ribo-Tri at 200 steps. A more detailed discussion of this tradeoff would help clarify the practical benefit of the additional architectural complexity.
Questions for authors
  1. What are the key methodological innovations of La-Ribo beyond the partially latent flow matching framework introduced by La-Proteina?
  2. Can the authors provide downstream generation results comparing different scaffold representations, rather than only reconstruction accuracy?
  3. Can the authors provide parameter-matched comparisons for La-Ribo-Base and La-Ribo-Tri, and discuss why triangle updates do not consistently improve codesignability?
5 comments
AuthorsOpenReview ↗

Response to Reviewer Ub79 (1/3)

We thank the reviewer for the careful reading. We answer the three questions in order, with new experiments for the second and third:

Q1 (novelty). La-Ribo builds on the recipe of La-Proteina, but RNA breaks one of its assumptions: the frame defined by the informative anchors (P, C4′, N) differs from the frame that atom construction needs. Our scaffold with a learned scaffold-to-sugar rotation addresses this, and we add a corpus for scarce RNA structures.

Q2 (scaffold representations). We retrained the autoencoder and the flow model for five alternative designs. Freezing the rotation, using a rigid ribose frame, or using fewer or different anchors all lower designability and codesignability, so both the P/C4′/N scaffold and the learned rotation contribute.

Q3 (Base vs. Tri). With parameter-matched variants, the gain of Tri does not come from parameter count. It lies in topology and diversity, while codesignability is on par with Base, and we explain why.

Q1. Contributions beyond La-Proteina.

We agree that La-Ribo builds on the recipe introduced by La-Proteina (a partially latent state and two-stage training with an autoencoder followed by joint flow matching). Applying this recipe to RNA is, however, not a direct transfer: the explicit state and its mapping back to atoms have to be redesigned, and the more direct adaptations we tested perform substantially worse (Q2). Beyond the recipe, our contributions are:

  1. A modeling challenge specific to RNA. La-Proteina keeps coarse coordinates explicit and absorbs the rest into latents. For RNA, the informative anchors, P (backbone), C4′ (sugar) and the glycosidic N (base attachment), define a frame RipR^{p}_i that differs from the ribose frame Ric=F(xiC4′,xiO4′,xiC3′)R^{c}_i = F(x^{\mathrm{C4'}}_i, x^{\mathrm{O4'}}_i, x^{\mathrm{C3'}}_i) that our atom builder needs (Sec. 3.2). The rotation between them varies across residues: it deviates from the global mean by 9.3∘9.3^{\circ} on average, and freezing it raises the frame error from 1.78∘1.78^{\circ} to about 14∘14^{\circ} (see Q2).
  2. A P/C4′/N scaffold with a learned scaffold-to-sugar rotation. The scaffold specifies the positions of P and the glycosidic N, which the ribose atoms O4′ and C3′ do not, and the decoder predicts a rotation QiQ_i with Ridec=RipQiR^{\mathrm{dec}}_i = R^{p}_i Q_i that maps the anchor frame to the sugar frame. Both are necessary (see Q2): within our framework, a single C4′ anchor, the RNA analogue of the Cα-only explicit state of La-Proteina, gives 24.28 designability, the rigid ribose frame with no transformation gives 27.94, and a frozen rotation gives 32.28, against 38.22 for La-Ribo-Base. For RNA, we observe that when the anchor frame differs from the decoder frame, learning the transformation between them appears to be what makes the anchors usable.
  3. A corpus that addresses data scarcity. We built 168,561 structures from experimental data and three folding models, including 10,631 new MSA-supported structures. Expanding the flow training data raises designability from 30.72 to 39.11 for La-Ribo and from 20.50 to 27.61 for RiboFlow, so the benefit is not specific to our representation (Tables 1 and 2).
AuthorsOpenReview ↗

Response to Reviewer Ub79 (2/3)

Q2. Downstream comparison of scaffold representations.

We retrained the autoencoder (encoder and decoder) and the flow model for each design below, with all other settings identical to Table 1 (data, optimizer, sampling schedule, 200 sampling steps). To finish within the rebuttal period we used the lighter Base architecture (no triangle updates) and scored with RhoFold. Metrics follow Table 1: strict [email protected] and [email protected], mean ± SD over three 600-structure sampling campaigns.

The decoder maps the scaffold pseudo-frame to the canonical sugar frame used by the atom builder, Ridec=RipQiR^{\mathrm{dec}}_i = R^{p}_i Q_i (Eq. 4), where Rip=F(xiC4′,xiνi,xiP)R^{p}_i = F(x^{\mathrm{C4'}}_i, x^{\nu_i}_i, x^{\mathrm{P}}_i) and Ric=F(xiC4′,xiO4′,xiC3′)R^{c}_i = F(x^{\mathrm{C4'}}_i, x^{\mathrm{O4'}}_i, x^{\mathrm{C3'}}_i). The variants are:

  • Fixed Q. Same P/C4′/N anchors, but QQ is no longer predicted. For every training residue we compute the target Qi∗=(Rip)⊤RicQ^{*}_i = (R^{p}_i)^{\top} R^{c}_i from the ground-truth frames, average these matrices over all N=12.3MN = 12.3\mathrm{M} training residues, and project the mean onto SO(3)SO(3) by SVD:

    Qˉ=ΠSO(3)(1N∑i=1NQi∗).\bar{Q} = \Pi_{SO(3)}\left(\frac{1}{N}\sum_{i=1}^{N} Q^{*}_i\right).

    This gives one constant rotation shared by all residues. The per-residue targets deviate from Qˉ\bar{Q} by 9.3∘9.3^{\circ} on average (median 5.4∘5.4^{\circ}), so a constant QQ is a reasonable first approximation but leaves residue-level frame error.

  • Identity Q (C4′/O4′/C3′). The scaffold anchors are C4′, O4′ and C3′, which span the canonical frame RicR^{c}_i that the atom builder uses. The frame built from the scaffold anchors therefore equals RicR^{c}_i and Qi=IQ_i = I: no transformation is needed. This is the rigid-frame scaffold of frame-based generators such as RNA-FrameFlow. It contains no phosphate or base-attachment anchor.

  • C4′ only, P/C4′, P/C4′/C3′ (learned Q). Fewer or different anchors, with QQ learned as in our model.

RhoFold, 200 steps

ScaffoldQDesign. [email protected]↑Design. [email protected]↑Codesign. [email protected]↑Codesign. [email protected]↑
La-Ribo-Baselearned38.22±1.2753.00±0.6732.56±2.2556.72±1.78
P/C4′/Nfixed32.28±2.4450.44±1.8027.17±0.7651.72±1.07
C4′/O4′/C3′identity27.94±0.6741.00±0.7329.94±1.7150.83±2.84
C4′learned24.28±0.7540.22±2.4125.17±1.7347.11±0.84
P/C4′learned22.44±0.8435.83±2.9625.11±0.3542.67±1.80
P/C4′/C3′learned31.83±3.1947.11±2.2530.94±1.7553.72±0.96

The first row is the original model of Table 1; the other rows are retrained variants.

(1) Learning QQ matters. Freezing QQ lowers designability from 38.22 to 32.28 and codesignability from 32.56 to 27.17. The decoder also becomes less accurate: with a constant QQ the frame error is 14.2∘14.2^{\circ} and the non-scaffold RMSD 0.330 Å, against 1.78∘1.78^{\circ} and 0.087 Å for the learned QQ (Table 10).

(2) Ribose-only anchors give lower designability. The identity-QQ variant reaches 27.94 designability ([email protected] 41.00) against 38.22 (53.00), and its codesignability is also lower (29.94 vs 32.56). O4′ and C3′ describe local ribose geometry but do not specify the positions of P and the glycosidic N, which carry the backbone trace and the base attachment; this variant removes P and N. This is consistent with the loss of that information, even though no QQ needs to be learned. Mapping a rigid-frame scaffold directly onto RNA is therefore less suitable than our P/C4′/N scaffold with a learned QQ.

(3) The other designs are also weaker. C4′ only, P/C4′ and P/C4′/C3′ reach 22.44 to 31.83 designability against 38.22, and 25.11 to 30.94 codesignability against 32.56. Replacing the base-attachment anchor N by the ribose atom C3′ costs several points of designability, which supports that N carries information beyond the sugar.

Together, in this setting (Base architecture, 200 steps, RhoFold), removing the learned QQ, removing P and N, or using fewer anchors each lowers designability and codesignability, so both the P/C4′/N anchors and the learned transformation contribute to the reported performance.

AuthorsOpenReview ↗

Response to Reviewer Ub79 (3/3)

Q3. Parameter-matched Base vs. Tri, and the codesignability trade-off

In brief: the gain of Tri does not come from parameter count. It lies at the structure level (topology, diversity, geometry), while strict codesignability is on par with Base.

1. Parameter-matched experiment. La-Ribo-Base (58.82M) and La-Ribo-Tri (65.41M) differ by about 11% in parameters. We trained two variants that change only the token width of the flow network, and evaluated them with the same sampling settings as the paper's models: La-Ribo-Tri-Slim (58.52M, fewer parameters than Base) and La-Ribo-Base-Wide (66.05M, more than Tri).

  • La-Ribo-Tri-Slim: token width 480, 58.52M, slightly fewer parameters than Base.
  • La-Ribo-Base-Wide: token width 544, 66.05M, slightly more parameters than Tri.

RhoFold, 200 steps

ModelParamsDesign. [email protected]↑Design. [email protected]↑Codesign. [email protected]↑Codesign. [email protected]↑
La-Ribo-Base58.82M38.22±1.2753.00±0.6732.56±2.2556.72±1.78
La-Ribo-Tri65.41M39.11±2.1258.83±1.4832.06±0.7557.89±1.71
La-Ribo-Tri-Slim58.52M37.39±0.9255.89±2.4129.17±0.6750.67±3.00
La-Ribo-Base-Wide66.05M32.61±0.1048.72±0.7925.67±1.0149.72±0.98

Mean ± SD over three 600-structure sampling campaigns; paper rows from Tables 1 and 12.

2. Not a parameter effect. Tri-Slim has fewer parameters than Base, yet its designability (37.39) and TM (55.89) are close to those of Tri (39.11, 58.83), and its TM exceeds that of Base (53.00). Base-Wide has more parameters but does not improve. At matched size, Tri-Slim against Base gives designability −0.8 and TM +2.9, and Tri against Base-Wide gives +6.5 and +10.1.

3. What Tri improves. [email protected] is higher (58.83 vs 53.00, Table 1), Tri yields more clusters than Base in all 8 paper comparisons (and Tri-Slim 28.0 vs 23.0 for Base), and it has 37% fewer clashes (Table 16).

4. Why codesignability does not improve uniformly. The two are on par in the paper (32.06±0.75 vs 32.56±2.25). Codesignability scores the sequence the model itself emits, whereas designability scores eight redesigned sequences for the generated backbone. Triangle updates act on the pair features and mainly improve backbone geometry, and their effect on the latents that carry sequence information may be more indirect. Table 3 shows that the update frequency matters: with updates every 4 or 3 blocks, codesignability drops to 25.17 and 26.83, and it returns to 32.06 only with updates every 2 blocks (Tri).

5. Practical benefit. Tri is preferable when backbone quality, topology and diversity matter most; Base is sufficient when codesignability and sampling speed are the priority.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to Reviewers and AC

We thank the Area Chair and Reviewers Ub79, ikdr, and EwQ6 for their constructive feedback, and summarize our new experiments and clarifications below:

1. Contributions beyond La-Proteina and downstream ablations (Ub79, EwQ6)

La-Ribo builds on La-Proteina's partially latent flow-matching recipe, but RNA is not a direct transfer: La-Proteina keeps one Cα explicit per residue, whereas RNA's P (backbone), C4′ (sugar) and glycosidic N (base) define a frame distinct from the canonical ribose frame for atom building, with a residue-varying rotation between them (9.3° on average from its global mean, median 5.4°). Beyond the recipe, we contribute (a) this P/C4′/N scaffold with a learned per-residue rotation Q, and (b) a 168K-structure corpus (item 3). We retrained the autoencoder and flow model for five variants as in Table 1 (Base, RhoFold, 200 steps; %, mean ± SD, three 600-structure campaigns):

AnchorsQDes. [email protected]↑Des. [email protected]↑Codes. [email protected]↑Codes. [email protected]↑
P/C4′/N (La-Ribo-Base)learned38.22±1.2753.00±0.6732.56±2.2556.72±1.78
P/C4′/Nfixed32.28±2.4450.44±1.8027.17±0.7651.72±1.07
C4′/O4′/C3′ (rigid ribose)identity27.94±0.6741.00±0.7329.94±1.7150.83±2.84
C4′ onlylearned24.28±0.7540.22±2.4125.17±1.7347.11±0.84
P/C4′learned22.44±0.8435.83±2.9625.11±0.3542.67±1.80
P/C4′/C3′learned31.83±3.1947.11±2.2530.94±1.7553.72±0.96

Learned P/C4′/N has the highest mean in all four metrics, while direct transfers perform much worse: C4′ alone (La-Proteina's Cα-only analogue) gives 24.28% strict designability, and the rigid ribose frame without P and N gives 27.94%. Fixing Q lowers designability (38.22% to 32.28%) and codesignability (32.56% to 27.17%), and raises frame error (1.78° to 14.2°) and non-scaffold RMSD (0.087 to 0.330 Å, Table 10); replacing N by C3′ gives 31.83%. When the anchor and atom-building frames differ, learning Q makes the anchors usable. (Ub79, 1/3–2/3; EwQ6, 1/2)

2. Triangle updates at similar parameter counts (Ub79, ikdr)

Base (58.82M) and Tri (65.41M) differ by 11% in size. We trained two width variants, Tri-Slim (Tri, width 480, 58.52M) and Base-Wide (Base, width 544, 66.05M), under the same 200-step RhoFold setup:

ModelParamsDes. [email protected]↑Des. [email protected]↑Codes. [email protected]↑Codes. [email protected]↑
La-Ribo-Base58.82M38.22±1.2753.00±0.6732.56±2.2556.72±1.78
La-Ribo-Tri-Slim58.52M37.39±0.9255.89±2.4129.17±0.6750.67±3.00
La-Ribo-Tri65.41M39.11±2.1258.83±1.4832.06±0.7557.89±1.71
La-Ribo-Base-Wide66.05M32.61±0.1048.72±0.7925.67±1.0149.72±0.98

Tri-Slim is smaller than Base yet close to Tri on designability and above Base on [email protected] (+2.9), while Base-Wide is larger than Tri yet below Base (Tri vs. Base-Wide: +6.5 strict designability, +10.1 [email protected]), so Tri's structure-level gains are not from parameter count. Tri-Slim also reaches 28.0 vs. 23.0 clusters for Base, and Tri has fewer clashes than Base (Table 16). Strict codesignability is on par (32.06±0.75% vs. 32.56±2.25%): triangle updates refine pair features and backbone geometry, affecting sequence latents more indirectly (updating every 4 or 3 blocks instead of 2 lowers codesignability to 25.17% and 26.83%, Table 3). Tri is preferable for backbone quality, topology and diversity; Base is sufficient when codesignability and speed are the priority.

On diversity, Tri exceeds Base in clusters across every paper comparison. At 200 steps (designability), cluster counts for Tri, Base and RNA-FrameFlow are 27.67, 23.00 and 26.00 with RhoFold, and 39.67, 31.33 and 35.00 with OpenDDE, so Tri has 20% (RhoFold) and 27% (OpenDDE) more clusters than Base. Counts cover refolds passing scTM ≥ 0.45; backbones are identical, while passing sets differ by refolder. (Ub79, 3/3; ikdr)

3. Training corpus and evaluation scope (Ub79, ikdr, EwQ6)

Our 168,561-structure corpus combines experimental data and three predictors (OpenFold3, NuFold, trRosettaRNA2), including 10,631 new MSA-supported structures. Expanding flow-training data raises strict designability from 30.72% to 39.11% for La-Ribo-Tri and from 20.50% to 27.61% for RiboFlow (Tables 1–2), so the gain is not representation-specific (Ub79, 1/3). For bpRNA-1m(90)/RNAStralign (currently trRosettaRNA2), we plan to release and compare structures from predictors such as Protenix. The dataset retains 47 of 48 NuFold species (list in EwQ6, 2/2). Designability and codesignability measure in-silico sequence–structure self-consistency; we will state this in Limitations and hope to report wet-lab validation in future work (ikdr).

We welcome further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.