acceptodds
Under review as a conference paper at ICLR 2027

REFRACT: Generating RNA Conformational Ensembles with Context-Aware Flows

Abstract

Generating RNA conformational ensembles that reflect their molecular environment remains challenging. We introduce , an -equivariant framework that couples flow-based reconstruction with context-aware transport. During training, the model learns structural regularities around observed conformations by reconstructing individual experimental structures from corrupted inputs, without requiring paired alternative conformations or molecular dynamics (MD) trajectories. CrossIPA incorporates fixed protein geometry through geometric cross-attention, conditioning RNA updates on the evolving RNA–protein spatial relationships. At inference, context-aware transport uses these learned updates to guide independently corrupted RNA conformations toward ensembles compatible with the supplied protein context. On protein-context ensemble recovery and conformational transition across binding conditions, achieves average matched TM-scores of and , respectively, exceeding the evaluated baselines. It also more closely matches nucleotide-level fluctuations and conformational distributions in MD reference ensembles. An ablation on the protein-bound benchmark shows that supplying protein context improves conformational recovery and reduces the proportion of generated structures with steric clashes.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
RNA ensemble generation, flow matching model

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ufWwOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

Given a sequence, an experimental starting structure and optionally a fixed protein context, this paper generates an RNA conformational ensemble with an SE(3) frame model trained by flow matching to reconstruct experimental structures from corrupted inputs; no paired conformations or MD trajectories are needed. CrossIPA lets RNA nucleotides attend jointly to RNA and protein residues, re-evaluating the RNA-protein geometry at each update, and at inference independently corrupted copies of the input are transported by the same model. REFRACT reports matched TM-scores of 0.669 on protein-bound recovery and 0.648 on binding-condition transition, closer MD agreement than RNAbpFlow and DynaRNA, and an ablation that raises AMR-R from 0.643 to 0.669 while lowering clashing structures from 32.7% to 14.1%.

Critical strengths
  1. The paper conditions RNA ensemble generation on bound protein, which the evaluated ensemble generators do not do. The authors state that "neither RNAbpFlow nor DynaRNA explicitly conditions exploration on bound protein structures" (Section 1), and CrossIPA recomputes the RNA-protein geometry at each update instead of treating the protein as a fixed feature. Its consequence for my assessment is that the paper addresses a setting the two closest methods do not cover, so the results are not a re-measurement of an existing benchmark.

  2. The evaluation is complete. Three complementary benchmarks come with an input-structure control, a local-perturbation control and cluster-level bootstrap intervals, and Table 6 reports all 20 RMSF profiles including the failures (2VPL free, MAE 4.61 A). The paper reports both improvements and failures, which lets me judge where the method works and where it does not.

Critical weaknesses
  1. The evidence does not isolate the proposed mechanism. The protein ablation removes context only at inference, using the same trained model, and it moves AMR-R by 0.026 (0.643 to 0.669). The same model without protein at inference still exceeds the strongest baseline by 0.258 (0.643 vs 0.385). The paper does not retrain a model without protein or without CrossIPA. It therefore shows a small benefit from inference-time context; it does not show that the architecture explains the gap to baselines.

  2. The setup supplies a correct experimental starting structure and a correctly placed protein. Section 2 states that the protein "is supplied with a known position and orientation relative to the starting RNA", and transport starts from 20-50% corruption of that structure (training uses t/T <= 0.5). The paper does not test three settings: no starting structure (de novo), a predicted rather than experimental start, and a protein-side holdout. The demonstrated capability is sampling around a known state.

Questions for authors
  1. At inference, sample 100 structures from the protein-free variant, discard those that clash with the protein, and report AMR-R and COV-R at the remaining budget. Does CrossIPA add more than this clash filter?

  2. Retrain without protein context and without CrossIPA, and report Tables 1-3. The current ablation changes only inference, so it cannot separate training-time protein exposure from inference-time conditioning.

  3. Did every method use the same candidate budget? AMR-R and COV-R take a maximum over candidates and increase with N, and Section 3.1 specifies N = 100 only for REFRACT. Please state N for AF3 subMSA, RhoFold+ subMSA, RNAbpFlow and DynaRNA.

  4. What base-pair input did RNAbpFlow receive, experimental or predicted? And why are AF3 and RhoFold+ absent from the MD comparison (Table 3) when they serve as ensemble baselines in Tables 1-2?

  5. Replace the experimental starting structure with an AF3 or RhoFold+ prediction and report the change. Without this, the method cannot be used where no experimental structure exists.

  6. For RBPTrans, is the target protein placed using the experimental coordinates of the target complex? If so, the benchmark supplies the relative RNA-protein placement that an application would need to determine, and the setup should state this.

2 comments
AuthorsOpenReview ↗

Overview & Response to Q1–Q3

We thank you for your constructive review. We ran all requested experiments (Q1–Q6), which address your two weaknesses:

  1. W1 — Isolating the mechanism (Q1–Q2): Retraining without protein and without CrossIPA separates our two contributions. (i) Structure-conditioned partial flow matching drives the global recovery lead over sequence-only baselines (0.664 vs. 0.385 AMR-R on RBP). (ii) CrossIPA itself drives performance wherever the protein matters: it cuts clashes (14.1% vs. 20.6–27.2%), raises partner-transfer coverage ([email protected] 0.853 vs. 0.735–0.779) and improves bound-state MD (W2W_2 2.85 vs. 3.25–3.27 Å).
  2. W2 — Scope and input assumptions (Q5–Q6): Initializing REFRACT from an AF3 prediction improves on that prediction (0.338 → 0.369 on RBP; 0.347 → 0.418 on RBPTrans). Replacing the target-complex protein placement with only the source complex (no target-state information) leaves recovery unchanged (0.67, 0.71) and does not increase clashes (10.6%, 9.8%). We will frame REFRACT as a structure-conditioned ensemble generator around a known or high-confidence structure.

Response to Q1 & W1 (Clash filtering vs. CrossIPA)

On RBP (41 clusters) we sampled N=100N=100 structures per cluster from the protein-free variants, discarded those clashing with the protein (any C4′ within 1 Å of a protein heavy atom), and scored the rest:

Table R3-1: Post-hoc clash filtering vs. CrossIPA on RBP.

VariantClash-free samples / 100AMR-R↑
REFRACT (unfiltered)100 (86.2 clash-free)0.669
REFRACT + clash filter86.20.668
No protein at inference + clash filter68.10.621
Retrained without protein + clash filter73.40.663

At the same generation budget (100 trajectories), CrossIPA yields 1.17–1.27× as many sterically valid conformations (86.2 vs. 68.1 and 73.4) and higher ensemble recovery (0.668 vs. 0.621 and 0.663): steering trajectories away from occluded space during transport avoids wasting samples on protein clashes.

Response to Q2 & W1 (Retraining without protein and without CrossIPA)

We retrained two ablated models from scratch with identical hyperparameters: (1) without protein, which never sees protein in training or inference; (2) without CrossIPA, where the protein enters only as a static per-nucleotide RBF distance feature (not updated across blocks via geometric cross-attention). We evaluate on RBP, on the RBPTrans transfer task (34 source→target pairs, the protocol of Fig. 4) and on MD:

Table R3-2: Retrained ablations. Transfer: AMR-R↑ / [email protected]↑ / [email protected]↑. MD: Ens. RMSD↓ / W2W_2↓ (Å).

ModelRBP AMR-R↑RBP Clash↓TransferMD boundMD free
REFRACT0.66914.1%0.648 / 0.853 / 0.4712.06 / 2.852.80 / 4.36
Without CrossIPA (static feat.)0.66020.6%0.628 / 0.779 / 0.3242.30 / 3.252.86 / 4.39
Without protein0.66427.2%0.623 / 0.735 / 0.3822.27 / 3.272.89 / 4.47
  1. Partial flow matching explains the gap to baselines: the model without protein already reaches 0.664 AMR-R on RBP (vs. 0.385 for the strongest sequence-only baseline), because the experimental states within a cluster share most of their global fold.
  2. CrossIPA improves the generated ensemble wherever the protein matters: it halves the clash rate (14.1% vs. 27.2%); on transfer, where the RNA must adapt to a different partner, it raises target-state coverage from 0.735 to 0.853 ([email protected]) and from 0.382 to 0.471 ([email protected]); on MD it improves the bound ensembles (Ens. RMSD 2.06 vs. 2.27 Å; W2W_2 2.85 vs. 3.27 Å) and leaves the free ensembles unchanged, as expected.
  3. The gain comes from CrossIPA itself, not from mere exposure to the protein: the static-feature variant stays close to the protein-free model on transfer and bound MD, and recovers only part of the clash reduction (20.6%).

We will add Table R3-2 to the paper and state the two contributions separately.

Response to Q3 (Candidate budget NN)

REFRACT, AF3 subMSA, RNAbpFlow and DynaRNA all generate N=100N=100 structures per cluster. RhoFold+ is deterministic, so its diversity comes only from MSA subsampling; with a singleton MSA (46/68 clusters in the default setup) it yields N=1N=1 (mean N=22.7N=22.7 on RBP). Table R3-3 compares N=1N=1 (mean of 20 random draws) with N=100N=100:

Table R3-3: AMR-R↑ at N=1N=1 vs. N=100N=100.

MethodRBP, N=1N=1RBP, N=100N=100RNA-only, N=1N=1RNA-only, N=100N=100
REFRACT0.5470.6690.7170.815
AF3 subMSA0.3280.3810.5280.629
RNAbpFlow0.3030.385—0.397
RhoFold+ subMSA0.3420.349—0.365

Even with a single candidate, REFRACT (0.547 on RBP, 0.717 on RNA-only) exceeds every baseline at N=100N=100. Giving RhoFold+ deeper MSAs from the AF3 pipeline raises its mean NN on RBP from 22.7 to 51.7 and its AMR-R from 0.349 to 0.364, leaving all conclusions unchanged. We will report NN for every baseline in Section 3.1.

AuthorsOpenReview ↗

Response to Q4–Q6

Response to Q4 (RNAbpFlow base-pair inputs and adding AF3/RhoFold+ to the MD benchmark).

  • RNAbpFlow inputs: It received predicted (not experimental) secondary structures—specifically, the ensemble of three base-pair maps predicted by RibonanzaNet, SPOT-RNA, and IPknot, matching its original publication protocol.
  • AF3 and RhoFold+ on MD: They were initially omitted from Table 3 because sequence-only predictors cannot distinguish between the free and protein-bound simulation conditions of the same RNA. We have now added both to the MD benchmark:

Table R3-4: MD benchmark including AF3 subMSA and RhoFold+ subMSA. (RhoFold+ produces >1>1 structure on only part of the conditions and is evaluated on that subset.)

Evaluation SubsetMethodRMSF error (Å) ↓RMSF rr ↑Ens. RMSD (Å) ↓W2W_2 (Å) ↓
All conditionsREFRACT1.311.310.670.672.472.473.723.72
All conditionsAF3 subMSA1.681.680.570.574.364.365.925.92
All conditionsRNAbpFlow3.543.540.540.544.634.637.117.11
All conditionsDynaRNA6.966.960.250.258.158.1511.4911.49
Conditions w/ RhoFold+ ensembleREFRACT1.231.230.630.632.682.683.803.80
Conditions w/ RhoFold+ ensembleAF3 subMSA1.851.850.500.503.583.584.994.99
Conditions w/ RhoFold+ ensembleRhoFold+ subMSA2.032.030.280.283.113.114.194.19

AF3 subMSA is the strongest sequence-only baseline on MD, but REFRACT outperforms both AF3 and RhoFold+ across all four metrics on both subsets.

Response to Q5 & W2 (Initializing REFRACT from AF3-predicted structures and scope of the method).

We replaced the experimental starting structure AA with the top-ranked AF3 prediction (AF3 top-1, full MSA), superposed onto AA with AA's native protein context:

Table R3-5: Ensemble recovery (AMR-R ↑\uparrow) initialized from predicted vs. experimental structures.

Initialization / MethodRBPRBPTransRNA-only
AF30.3380.3380.3470.3470.5440.544
REFRACT initialized from AF30.3690.3690.4180.4180.5640.564
REFRACT initialized from experimental AA (paper)0.6690.6690.6950.6950.8150.815

REFRACT consistently improves upon its starting AF3 prediction across all three benchmarks (0.338→0.3690.338 \to 0.369 on RBP; 0.347→0.4180.347 \to 0.418 on RBPTrans; 0.544→0.5640.544 \to 0.564 on RNA-only), showing that flow transport explores useful conformational variations even around a predicted start. At the same time, because single AF3 predictions on these post-cutoff complexes often have inaccurate global folds (mean TM-score=0.345\text{TM-score} = 0.345 to experimental AA on RBP), partial noising (ts/T≤0.5t_s/T \le 0.5) is not designed to correct a globally misfolded topology. Accordingly, REFRACT is positioned as a structure-conditioned ensemble generator (like Str2Str and DynaRNA) that samples conformational landscapes around an experimental structure or an accurate structural model. We will include Table R3-5 and state this scope clearly in the paper.

Response to Q6 & W2 (RBPTrans protein registration and target-free placement controls).

In the paper's RBPTrans setup, the source RNA was superposed onto the target RNA to place the target protein in its experimental pose (which we will state explicitly in Section 3.1). To test whether our results depend on this target-complex registration, we reran inference using only the source complex (AA paired with its own native protein in its source coordinates, carrying zero information from the target complex), as well as under rigid-body protein perturbations:

Table R3-6: Target-free protein placement controls.

Protein Input SettingRBP AMR-R ↑RBP Clash ↓RBPTrans AMR-R ↑RBPTrans Clash ↓
Target complex (paper)0.670.6714.1%14.1\%0.700.7013.5%13.5\%
Source complex only (zero target info)0.670.6710.6%10.6\%0.710.719.8%9.8\%
No protein (C=∅\mathcal{C} = \emptyset)0.640.6432.7%32.7\%0.680.6837.3%37.3\%

Using only the known source complex leaves recovery unchanged (0.670.67 and 0.710.71) and does not increase clash rates (10.6%10.6\% and 9.8%9.8\%) (and rotating the protein by 15∘–45∘15^\circ\text{--}45^\circ likewise preserves recovery while increasing clashes to 17.1%–23.4%17.1\%\text{--}23.4\% on RBP). This confirms that the reported gains do not depend on oracle target-complex registration. In practice, protein placement can come directly from the starting complex, or from rigid-body docking / co-folding when modeling a new partner.

Reviewer_5rt6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

REFRACT generates RNA conformational ensembles from a single experimental structure, optionally conditioned on a fixed protein partner. It uses the RNA-FrameFlow representation (C4′-centred frames + torsions), trains a Str2Str-style "DenoisingIPA" network on single structures (RNAsolo 2.0 + PRI30k, cutoff Jan 2024) with partial noising t/T ~ U(0.01, 0.5) and a C4′–Cα interface loss, and adds "CrossIPA" so RNA queries attend over fixed protein keys/values (zero pair bias for protein keys). At inference, "context-aware transport" noises the input to t_s/T ∈ {0.2–0.5} and integrates back to t=0.01T; 25 samples per level, pooled to N=100.

Results: RBP AMR-R 0.669 vs 0.385 best baseline, clash 14.1% vs ~80%. RNA-only AMR-R 0.815 vs 0.629 for AF3+MSA subsampling. RBPTrans AMR-R 0.648 vs 0.437. MD: best on all six metrics. Removing the protein at inference drops AMR-R 0.669→0.643 and raises clashes 14.1%→32.7%.

Critical strengths
  1. Protein-context ablation. The paper is the first in this line (vs. RNAbpFlow, DynaRNA) to model RNA ensembles as depending on a fixed protein environment, and it isolates the effect cleanly with one trained model, changing only the inference-time context: RBP clash 32.7%→14.1% at similar diversity (2.59 vs 2.52 Å); RBPTrans [email protected] 0.588→0.853. This is the cleanest evidence that CrossIPA conditioning does something real, which matters since "context-aware" is the paper's main pitch.
  2. Careful metrics and honesty about limits. App. B defines AMR-R, COV-R, clash, RMSF error, W2, PC-sim precisely and uses cluster-level bootstrap CIs (pairing same-RNA entries in RBPTrans), while admitting the CIs don't capture seed variance. App. E openly flags the Fig. 8 profiles as cherry-picked "lowest-MAE" cases and still reports all 20 in Table 6. This transparency makes the paper auditable and partly offsets small benchmark sizes.
Critical weaknesses

The biggest problem is possible input leakage, with trivial controls missing. REFRACT (and DynaRNA) start from an experimental structure of the same RNA, while AF3/RhoFold+/RNAbpFlow see only sequence (or sequence + base pairs). The paper never says whether the input A is itself one of the reference conformations scored by AMR-R/COV-R — if it is, recovering it at t_s/T=0.2 is close to guaranteed. Three things point this way: Sec. 3.1 implies multiple registered structure–context pairs may be supplied per cluster; Table 2's COV-R is identically 0.755 at both the 0.7 and 0.8 thresholds, which is what near-exact copies would produce; and App. B mentions an "input-structure control" that is never actually reported for Tables 1–2 or Fig. 4. RBPTrans needs the analogous control (source-vs-target TM-score) — without it 0.648 is uninterpretable, especially since the protein-free run already hits 0.602. There's a related leak risk in how the target protein's position/orientation is registered relative to the source RNA in RBPTrans: if this comes from superposing the target complex, it indirectly encodes the target RNA pose, which the interface loss (Eq. 3) is explicitly trained to reproduce — this could also explain the big [email protected] jump. Finally, sequence-only baselines, apparently including AF3, are never given the protein, so their ~80% clash rates and the "RhoFold+ stays flat across conditions" contrast in Fig. 5 are largely by construction. Together these could change both the size and the meaning of the headline gains.

Second, the MD benchmark likely has train/test overlap, and the structure-conditioned baselines are weak. The Jan-2024 cutoff and 80% CD-HIT filter apply only to the main experimental sets; the MD systems (1ASY, 1JBS, 1MMS, 1OOA, 1RKJ, 2R8S, 2VPL, 2ZM5, 3IEV, 7M5O, 1R2P) are all pre-2024 PDB entries very likely already in RNAsolo/PRI30k training data, "reserved for evaluation" covering only the trajectories. A denoiser trained to reconstruct exactly these structures will naturally produce tight ensembles around the starting state — which is also what short MD around a crystal structure rewards. That's consistent with the trivial "Input structure" control already being strong (Table 3: RMSD 2.67 vs REFRACT's 2.47, W2 4.06 vs 3.72; Table 5 bound: W2 3.26 vs 2.85), so REFRACT's MD edge may be "staying close to a memorized input" rather than learned flexibility. Meanwhile DynaRNA, the only external structure-conditioned baseline, performs below sequence-only methods on RBP (AMR-R 0.328; COV-R 0.000 on RNA-only), suggesting it wasn't tuned properly, and there's no RNA-FrameFlow/Str2Str-style baseline without CrossIPA, no AF3-with-protein, and no REFRACT ablation without protein/interface loss. Without these, CrossIPA's specific contribution can't be isolated, and given the method itself is fairly incremental, the paper's case rests almost entirely on an empirical story that these gaps weaken.

Questions for authors
  1. For each RBP/RNA-only cluster: which structure is input A, and is A among the scored references Y? Please report (a) an input-copy baseline (AMR-R/COV-R of A itself), (b) results with A excluded from Y, (c) pairwise TM-score spread among references per cluster.
  2. How is the target protein placed relative to the source RNA in RBPTrans? If registration uses target-complex RNA coordinates, please rerun with a registration that doesn't use the target RNA pose (e.g., docking-free placement from the source complex, or perturbed placements), and report the source-vs-target TM-score baseline.
  3. Was AF3 run with the protein chain included? If not, please add AF3 co-folding (with/without MSA subsampling) on RBP and RBPTrans — the natural context-aware baseline.
  4. Are the MD starting structures (or ≥80%-identity homologs) in the training set? Please report MD results after retraining with them removed, and define "Local perturb." and explain its consistently negative RMSF correlation (~−0.27).
  5. How were DynaRNA's settings and REFRACT's t_s grid chosen — with a validation split separate from the test clusters?
  6. Please give architecture/training details (blocks, parameters, filtered dataset sizes, integration steps K, per-sample cost vs AF3/MD) and multi-seed results for the protein ablation — 0.643 vs 0.669 looks small next to the overlapping CIs in Fig. 7a.
5 comments
AuthorsOpenReview ↗

Response to Weaknesses, Q1 & Q2

We thank you for your rigorous examination of our controls. We first address W1 and W2, then give point-by-point results (Q1–Q2 here; Q3–Q6 in Post 2).

Response to W1: No Input or Target-Pose Leakage

Three direct controls show that no input or pose leakage exists in our setup:

  1. The network never sees the clean AA; its actual input is severely degraded. AA is first corrupted (ts/T∈[0.2,0.5]t_s/T\in[0.2,0.5]). This Noised input has an AMR-R of only 0.28 on RBP and 0.29 on RBPTrans, and clashes with the protein in 71.9% and 83.0% of cases (Table R2-2). REFRACT reconstructs valid conformations from these states, raising AMR-R from 0.28 to 0.67 and cutting clashes fivefold.
  2. Cross-basin exploration when AA is excluded (Q1). With AA removed from Y\mathcal{Y}, REFRACT stays well ahead of all external baselines. On distant targets (TM(AA, target) <0.5), REFRACT (0.45 on RBP, 0.46 on RNA-only) beats the uncorrupted Copy-AA (0.41, 0.40), showing that transport moves away from AA toward alternative states.
  3. No target-pose leakage (Q2); protein-aware AF3 baseline (Q3). Using the source complex only (no target-state information) leaves recovery unchanged (0.67, 0.71) and does not increase clashes (Table R2-2); AF3 RNA–protein co-folding (Q3) confirms REFRACT's lead over context-aware baselines.

Response to W2: MD Retraining, Context-Aware Baselines, and Architectural Ablations

We ran all corresponding experiments (Q3–Q6): (1) retraining with all MD homologs removed (Q4, Table R2-4) shows that nucleotide-level flexibility capture is unaffected by training overlap (RMSF error 1.49 vs. 2.74 Å free, 0.88 vs. 1.87 Å bound; r=0.56r=0.56); (2) 5-seed sampling (Q6, Table R2-5) shows that the ablation gap is not sampling noise; (3) retraining without protein and without CrossIPA (Reviewer ufWw Q2, Table R3-2) shows that partial flow matching drives global fold exploration, while CrossIPA halves protein clashes (14.1% vs. 27.2%), raises partner-transfer coverage ([email protected] 0.853 vs. 0.735) and improves bound-state MD (W2W_2 2.85 vs. 3.27 Å).

Response to Q1 (Input AA, Copy-AA baseline, excluding AA, reference TM spread)

For every RBP and RNA-only cluster, AA is the first conformer in our data release (conformer_01) and is included in Y\mathcal{Y} in Tables 1–2 (Copy-AA is the "input-structure control" of App. B). In Table 2, [email protected] and @0.80 are both 0.755 because, of the 45 RNA-only targets, 34 reach best TM ≥0.80 and 11 have TM ≤0.694 (none in [0.70, 0.80)): the similarity of the targets to AA is bimodal, not a sign of copying.

Table R2-1 stratifies targets by TM(AA, target):

Table R2-1: AMR-R↑ with AA included / excluded from Y\mathcal{Y}.

MethodRBP <0.5RBP 0.5–0.7RBP ≥0.7RNA-only <0.5RNA-only 0.5–0.7RNA-only ≥0.7
Median TM(AA, target)0.440.610.800.400.560.92
Copy-AA0.60/0.410.75/0.600.90/0.820.70/0.400.79/0.590.95/0.92
Random perturbation0.43/0.330.57/0.490.78/0.730.51/0.340.75/0.570.85/0.83
REFRACT0.51/0.450.63/0.570.80/0.770.58/0.460.76/0.590.87/0.85
AF3 subMSA0.28/0.260.39/0.390.41/0.410.44/0.460.57/0.580.68/0.68
DynaRNA0.28/0.280.34/0.340.36/0.350.26/0.270.28/0.260.32/0.33

Copy-AA is a single structure with zero diversity; its score on Y∖{A}\mathcal{Y}\setminus\{A\} only reflects how similar the other experimental states already are to AA. With AA excluded, REFRACT beats both Copy-AA and Random perturbation on distant targets (TM <0.5; 68% of RBP targets improved) and outperforms the external baselines. We will report both protocols and Table R2-1 in the main text.

Response to Q2 (RBPTrans protein registration and noised-input evaluation)

In the paper's RBPTrans setup, the source RNA was superposed onto the target RNA to place the target protein in its experimental pose (we will state this in Section 3.1). To test whether recovery depends on this, we reran inference with only the source complex (AA with its own protein, no target-state information) and also evaluated the Noised input (AA after forward corruption, before denoising):

Table R2-2: Recovery and clash under target-free protein registration and noised inputs.

Protein input / controlRBP AMR-R↑RBP Clash↓RBPTrans AMR-R↑RBPTrans Clash↓
Target complex (paper)0.6714.1%0.7013.5%
Source complex only0.6710.6%0.719.8%
No protein0.6432.7%0.6837.3%
Copy-AA0.78—0.81—
Noised input (actual network input)0.2871.9%0.2983.0%

Using only the source complex preserves recovery and steric compatibility without any target-pose registration. The contrast between the Noised input (AMR-R 0.28–0.29, 72–83% clash) and REFRACT (0.67–0.71, 10–14%) confirms context-guided reconstruction from heavily corrupted inputs.

AuthorsOpenReview ↗

Response to Q3, Q4, Q5 & Q6

Response to Q3 (AF3 co-folding with protein chains)

In the paper, AF3 was run on the RNA sequence only. We added AF3 co-folding with the protein chains (no templates, to prevent test-structure leakage), both with full MSAs and with the MSA subsampling of the RNA-only baseline:

Table R2-3: Comparison with AF3 RNA–protein co-folding.

MethodRBP AMR-R↑RBP Clash↓RBPTrans AMR-R↑RBPTrans Clash↓
AF3 subMSA, RNA only (paper)0.3878.9%0.4482.8%
AF3 co-folding, full MSA0.4251.3%0.4383.3%
AF3 co-folding, subMSA0.4666.8%0.5170.8%
REFRACT0.6714.1%0.7013.5%

As expected, supplying the protein improves AF3's recovery on RBP (0.38 → 0.42 with full MSAs, 0.46 with subsampling) and lowers its clash rate (78.9% → 51.3–66.8%); on RBPTrans the gain appears with subsampling (0.44 → 0.51). REFRACT remains clearly superior on both metrics.

Response to Q4 (MD homology, retraining, and Local perturbation)

Under our ≥80% sequence-identity rule, 985 of the 23,356 training entries (4.2%) are homologous to an MD system. We retrained REFRACT from scratch with all 985 removed and re-evaluated the MD benchmark:

Table R2-4: MD evaluation after removing all MD systems and ≥80% homologs from training.

ConditionMethodRMSF error (Å)↓RMSF rr↑Ens. RMSD (Å)↓W2W_2 (Å)↓
FreeCopy-input2.74—2.984.59
FreeLocal perturbation2.14−0.274.585.83
FreeREFRACT (paper)1.750.632.804.36
FreeREFRACT (retrained)1.490.562.904.43
BoundCopy-input1.87—2.213.26
BoundLocal perturbation2.30−0.274.054.84
BoundREFRACT (paper)0.810.692.062.85
BoundREFRACT (retrained)0.880.562.373.34

After removing all homologs, nucleotide-level flexibility prediction remains strong: the retrained model still cuts RMSF error by nearly half relative to Copy-input (1.49 vs. 2.74 Å free; 0.88 vs. 1.87 Å bound) and keeps a solid RMSF correlation (r=0.56r=0.56). Ensemble RMSD and W2W_2 move closer to Copy-input, because microsecond MD stays near the starting crystal structure whereas generative sampling has a broader spread (still far closer than Local perturbation). REFRACT's ability to capture flexibility profiles is therefore learned, not memorized. We will include this evaluation in the paper.

Why Local perturbation has r≈−0.27r\approx-0.27: it adds i.i.d. isotropic Gaussian noise (σ=4.0\sigma=4.0 Å, selected on three held-out 1R2P force-field conditions) to every C4′ atom. In least-squares superposition, peripheral atoms have larger lever arms, so more of their noise is absorbed by the fitted rotation, which lowers their apparent RMSF (r=−0.46r=-0.46 with distance from the centroid); in MD, peripheral loops and termini fluctuate the most (r=+0.51r=+0.51).

Response to Q5 (DynaRNA settings and tst_s grid)

DynaRNA was run with the authors' released checkpoint and default inference protocol (tstep=800/1024t_\text{step}=800/1024). For REFRACT, ts/T∈{0.2,0.3,0.4,0.5}t_s/T\in\{0.2,0.3,0.4,0.5\} (25 samples per level) is an untuned, evenly spaced grid up to the maximum training corruption (t/T≤0.5t/T\le0.5).

Response to Q6 (Architecture, runtime, and multi-seed / retrained ablations)

  • Specifications & runtime: REFRACT has 34.9M parameters in 4 blocks (8-head RNA IPA, cs=256c_s=256, cz=128c_z=128; 2-layer sequence Transformer; CrossIPA cx=64c_x=64 over the 128 nearest protein residues). It was trained on 23,356 CD-HIT-filtered structures (22,661 with protein) for 96,360 steps (Adam, lr 10−410^{-4}, batch size 10). Inference uses K=60K=60 steps: 0.15–0.37 s per sample on one GPU, <1 min per 100-sample ensemble, vs. ~11 s per seed plus MSA search for AF3 and GPU-days for MD.
  • Multi-seed & retrained ablations: The CIs in Fig. 7a reflect variance between the 41 clusters, not sampling noise. Across 5 sampling seeds (Table R2-5), REFRACT outperforms the protein-free run in all 5 (s.d. ≤0.003) and more than halves the clash rate, which is lower in 36/40 RBP clusters (p<10−8p<10^{-8}, Wilcoxon signed-rank). We also retrained REFRACT without protein and without CrossIPA (Reviewer ufWw Q2, Table R3-2): CrossIPA drives clash reduction (14.1% vs. 27.2%), partner-transfer coverage ([email protected] 0.853 vs. 0.735) and bound-state MD accuracy (W2W_2 2.85 vs. 3.27 Å).

Table R2-5: Protein ablation across 5 sampling seeds (mean ± s.d.).

VariantRBP AMR-R↑RBP Clash↓RBPTrans AMR-R↑RBPTrans Clash↓
REFRACT0.668 ± 0.00014.0 ± 0.4%0.697 ± 0.00214.0 ± 1.1%
No protein at inference0.643 ± 0.00132.8 ± 0.3%0.678 ± 0.00338.0 ± 1.2%

Thanks for the additional controls—they address several of my concerns. Still, copying the input scores higher on the original recovery metrics, and excluding the input shows only modest gains on distant targets, with no gains over copying in the other groups. The source-complex control is helpful, but it is unclear whether it still tests transfer to the target partner. These results leave me unconvinced by the breadth of the ensemble-recovery claims, so I’ll keep my original score.

Reviewer_xDvgOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces Refract, a SE(3)-equivariant flow matching model that samples ensembles of RNA 3D structures. The method features a SE(3) flow matching framework that is conditioned on the presence of a protein partner. The paper outperforms existing methods at sampling ensembles that deal with 3D structure modeling.

Critical strengths
  • RNA backbone parameterization seems to be in order and captures the different degrees of freedom relevant for a RNA nucleotide.
  • The use of CrossIPA as the denoiser architecture is apt given that the model processes both proteins and RNA structures. Good design choice. The authors also use auxiliary protein-RNA-complex losses that help the model with the underlying objective of getting the ensemble in the presence of a protein partner.
  • Promising results when trained on RNA ensembles, outperforming existing methods dedicated to and repurposed for ensemble sampling. The authors also state improverd results in the presence of a protein partner.
Critical weaknesses

I think the application of ML for RNA ensemble modeling is mishandled by the community given the dynamic nature of RNA and the heterogenous nature of their ensmebles.

For instance, RNAsolo doesn't have backbones that look too different from each other, mostly just loop and stem movements. We're not seeing significant rearrangements of 3D backbones that mediate many RNA functions. Furthermore, according to Figure 6, the bound and unbound/free states of the P4-P6 domain do not look that different, and perhaps might only experience base-related atom movements. Therefore, training a generatrive model, which is already inherently mode-seeking, on data with limited conformational heterogeneity yields a model with even worse heterogeneity. Q1 below helps me understand that profile better, which I think should be honestly reflected in the main paper.

Additionally, the chocie of RNAbpFlow as a baseline is concerning. RNAbpFlow, which is a spin-off of RNAFrameFlow, is a de novo design model. Hoping to sample 3D diversity from it while pinning down the sequence and 2d secondary structure seems underhanded, and not in line with the original model's usecase.

I think RNA structure prediction is still not great and has huge room for improvement. RNAFrameFlow and RNAbpFlow, while great at learning plausible 3d local geometries, are not good enough to accurately understand the molecular breathing motions unless they use some sort of physics-based tool that provides energy values.

Questions for authors
  1. Can the authors please add any tables of information on how heterogenous or diverse the ensembles are, especially for RBP and the MD datasets used? If the spread is poor (eg: RMSF, angular profiles, etc), then there is no point training a generatribe model on top of it because it will yield no rare states, and you're mostly seeing movements around the same metastable basin. If the ultimate application of such a model is to capture the conformational diversity, the underlying training data used here would not be appropriate.

  2. It says baselines were used without retraining, using the original released checkpoints. Would the authors kindly retrain RhoFold on the same RNAsolo data as Refract and rerun MSA subsampling? Otherwise, it's an unfair comparison. Also, could the authors please elucidate the MSA-subsampling process for AF3? Was nhmmer used here? I'd optionally also recommend checking out RiboSeek [1] to get RNA MSAs.

[1] https://www.biorxiv.org/content/10.64898/2026.07.31.741718v1

2 comments
AuthorsOpenReview ↗

Response to Weaknesses

We thank you for your insightful review. We fully agree that conformational diversity must be evaluated rigorously beyond local base fluctuations, and we address your three core concerns (W1–W3) below.

Response to W1: Why REFRACT does not require multi-state training data and is not inherently mode-seeking.

Your concern is that training a mode-seeking generative model on static structures with limited conformational heterogeneity will collapse predictions into a single basin. We address this from both our training design and theoretical formulation:

  • 1. REFRACT does not depend on multi-conformation training data: Unlike methods that map an RNA sequence directly to a full conformational ensemble, REFRACT is trained via partial corruption and reconstruction (t/T≤0.5t/T \le 0.5) on individual structures in RNAsolo and PRI30k. The model learns local 3D geometric regularities and protein-induced spatial constraints (via CrossIPA) from single static snapshots, so it does not require training RNAs to have multiple experimentally resolved conformations. At inference, adding different levels of noise (ts/T∈[0.2,0.5]t_s/T \in [0.2, 0.5]) to the starting structure and reconstructing it under the target protein environment allows the model to explore diverse, sterically valid conformations (Figure 1).
  • 2. Flow matching learns continuous distributions rather than collapsing to a single mode: Theoretically, flow matching and diffusion models are formulated to approximate the full conditional data distribution by matching probability paths, rather than being "inherently mode-seeking" like point-estimate regression or reverse-KL objectives. While sequence-only baselines in Tables 1–2 struggle to recover alternative states from sequence alone, REFRACT combines multi-level noise initialization with protein-conditioned flow transport to sample distinct 3D conformations.

This captures genuine 3D backbone rearrangements rather than mere base motions: in Figure 6, although the global fold of P4–P6 is preserved, the free (1HR2) and Fab-bound (2R8S) references differ by a C4′\text{C4}' RMSD of 5.855.85 Å, with 37/15737/157 nucleotides displaced by >5>5 Å (up to 27.527.5 Å) after global superposition. On RBPTrans, supplying protein context boosts target-state coverage (COV-R\@0.5\text{COV-R\@0.5}) from 0.5880.588 to 0.8530.853. We will highlight this conceptual distinction and add displacement annotations to Figure 6.

Response to W2: Clarifying RNAbpFlow (ensemble predictor) vs. RNA-FrameFlow (de novo design).

We apologize for not distinguishing these two works clearly enough. First, we emphasize that our evaluated baseline RNAbpFlow (Tarafder & Bhattacharya, 2026) is specifically an RNA 3D ensemble prediction model, not a de novo design model—its standard inputs are a fixed RNA sequence and predicted 2D secondary structure (base-pair maps), and it formulates structure prediction as conditional 3D ensemble generation. By contrast, RNA-FrameFlow (Anand et al., 2025) is an unconditional de novo backbone design model, from which we only adopt the SE(3)SE(3) nucleotide frame representation. Evaluating RNAbpFlow under fixed sequence and predicted secondary structure is therefore completely aligned with its original use case. We will clarify this in Sections 3.1 and 4.

Response to W3: Addressing current limitations in RNA structure prediction and physical energy tools.

First, we fully agree that existing sequence-based RNA structure predictors still struggle to capture conformational dynamics—and this limitation is precisely what motivates REFRACT, which introduces a complementary paradigm that leverages a known starting structure and bound protein geometry to guide conformational exploration (and even improves upon AF3 starting predictions; Table R3-5). Second, while REFRACT (like RNAbpFlow) is currently purely data-driven and does not call an explicit physics-based energy tool, its learned velocity field conditioned on 3D RNA topology and protein contacts captures anisotropic collective motions rather than unguided local noise. On our MD benchmark (Table 3), isotropic Gaussian noise (Local perturbation) yields a negative per-nucleotide RMSF correlation (r=−0.274r = -0.274) and poor subspace agreement (PC-sim=0.069\text{PC-sim} = 0.069), whereas REFRACT achieves strong positive RMSF correlation (r=0.674r = 0.674), the lowest Wasserstein distance (W2=3.72W_2 = 3.72 Å), and the highest principal-component similarity (PC-sim=0.507\text{PC-sim} = 0.507). Integrating explicit physics-based energy terms into our flow velocity field is a promising direction for future work.

AuthorsOpenReview ↗

Response to Q1 & Q2

Response to Q1 (Ensemble heterogeneity and base-pair rearrangements)

REFRACT is trained only on single static structures, so our benchmarks are used strictly for evaluation: Benchmark I (RBP, RBPTrans, RNA-only) tests discrete cross-basin rearrangements; Benchmark II (MD) tests continuous intra-basin flexibility.

1. Benchmark I. Every cluster is backed by literature confirming distinct functional states. Table R1a quantifies the spread across reference states (C4′/P backbone atoms, barnaba annotations):

Table R1a: Heterogeneity of Benchmark I (medians).

BenchmarkMax per-nt C4′ disp. (Å)nts with Δ(η/θ)>60∘\Delta(\eta/\theta)>60^\circBase pairs remodeled
RBP (41 clusters)19.5 (≥5 Å in 39/41)12.3%40%
RBPTrans (34 pairs)10.1 (≥5 Å in 22/34)12.2%54%
RNA-only (17 clusters)14.7 (≥5 Å in 11/17)6.1%38%

In RBP, 18/41 clusters have inter-state C4′ RMSD ≥5 Å and 13/41 have TM-score <0.5. 38–54% of base pairs are remodeled; even among high-similarity pairs (TM ≥0.8), 35/41 break or form canonical WC/GU pairs.

To verify that REFRACT captures these rearrangements, we evaluate reproduction of the base pairs gained or lost between input AA and each target (Δbp Recall/F1), INF and eRMSD (DynaRNA outputs C4′ only and is omitted):

Table R1b: Reproduction of base-pair rearrangements (Δbp Rec↑/F1↑; INF↑/eRMSD↓).

MethodRBP ΔbpRBP INF/eRMSDRBPTrans ΔbpRBPTrans INF/eRMSDRNA-only ΔbpRNA-only INF/eRMSD
REFRACT0.747/0.4850.651/0.8420.715/0.5120.632/0.9550.677/0.4010.629/1.025
AF3 subMSA0.568/0.3330.495/1.2630.572/0.4150.605/1.1490.577/0.4690.720/1.028
RhoFold+ subMSA0.496/0.2120.164/2.0380.599/0.2790.166/1.2370.583/0.2080.153/3.447
RNAbpFlow0.694/0.3470.506/1.2530.635/0.4250.577/1.0870.672/0.3570.522/1.412

REFRACT leads on all four metrics on RBP and RBPTrans.

2. Benchmark II (MD). The MD ensembles (17.2 μs, 11 systems) show substantial heterogeneity: median per-nucleotide C4′ RMSF 1.81 Å (max 4.32 Å), intra-ensemble pairwise RMSD 2.68 Å, and 12% of nucleotides with η/θ\eta/\theta circular SD >30°. Across the 9 paired free/bound systems, cross-ensemble RMSD is 1.89× the within-ensemble RMSD (5.55 vs. 2.67 Å), a clear protein-induced shift. We will add both tables to the paper.

Response to Q2 (Retraining RhoFold+, RiboSeek MSAs, and AF3 MSA subsampling)

1. Retraining RhoFold+. RhoFold+ releases only pre-trained weights and inference code, without training scripts. To assess fairness, we computed the maximum sequence identity of our 68 benchmark RNAs to each training source:

Training sourceMedian max identityIdentity ≥0.9Identical
REFRACT training set0.600/680/68
RhoFold+ (BGSU 2022-04-13)0.7221/6813/68

REFRACT's training set excludes homologs (≤80% CD-HIT identity, Jan-2024 cutoff), whereas RhoFold+'s contains near-identical sequences for 21/68 benchmark RNAs, so the current comparison is favorable to RhoFold+.

2. RiboSeek MSAs. Following your suggestion, we generated MSAs with RiboSeek (Rfam, RNAcentral, NT-RNA), raising median MSA depth from 8 to 48 rows (singletons 26/68 → 16/68):

Table R2: Recovery with RiboSeek MSAs, AMR-R↑ ([email protected]↑).

MethodMSARBPRBPTransRNA-only
AF3 subMSAnhmmer (paper)0.381 (0.195)0.441 (0.444)0.629 (0.735)
AF3 subMSARiboSeek0.378 (0.183)0.436 (0.500)0.582 (0.618)
RhoFold+ subMSAPaper0.349 (0.232)0.475 (0.500)0.365 (0.294)
RhoFold+ subMSARiboSeek0.377 (0.256)0.486 (0.667)0.462 (0.353)
REFRACT—0.669 (0.817)0.695 (0.833)0.815 (0.951)

RiboSeek improves RhoFold+ (up to +0.10 on RNA-only) and leaves AF3 similar; REFRACT keeps a +0.29 AMR-R lead on RBP.

3. MSA-subsampling protocol. Both baselines follow del Alamo et al. (2022), using nhmmer MSAs from AF3's default RNA pipeline: the visible MSA is capped at 128 randomly drawn rows (down from 1024), with 1 recycle and 100 seeds (1 diffusion sample per seed). We will add these details and Table R2 to the Appendix.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.