acceptodds
Under review as a conference paper at ICLR 2027

Recurrent Geometric Coupling Refinement for RNA Inverse Folding

Abstract

RNA inverse folding seeks nucleotide sequences compatible with a target three-dimensional structure. Base-pairing and stacking interactions couple nucleotide choices in RNA, yet sampling positions independently within each generation step can fail to preserve their joint compatibility. We introduce **Recurrent Geometric Coupling Refinement (RGCR)**, which builds on **Masked Diffusion Language Modeling (MDLM)** to learn joint nucleotide distributions over geometry-derived pairs and triples. Estimates of their expected log-score gains over singleton predictions guide geometric message passing that recurrently refines both the joint distributions and these estimates. The final scores select disjoint groups for joint sampling among the positions generated at each step. We establish a normalized generative process and decompose its predictive gains into contributions from dependence retention and marginal prediction. RGCR improves structure fidelity and recovery performance on the RNA inverse-folding benchmark. Code and checkpoints are available at https://anonymous.4open.science/r/rgcr-5241/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Geometric Deep Learning, RNA Inverse Folding, Geometry Coupling, Dependence Retention

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_XZRaOpenReview ↗
Rating
3
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Recurrent Geometric Coupling Refinement (RGCR) for RNA inverse folding. The method extends masked diffusion generation with geometry-derived pair and triple nucleotide distributions. A recurrent refinement mechanism updates predictive beliefs and estimates the expected gains of joint predictions over independent singleton predictions. These learned gains are subsequently used to select disjoint groups for joint sampling.

The authors establish a normalized generation process, derive an exact likelihood conditioned on the reveal schedule, and analyze predictive gains in terms of dependence retention and prediction error.

On the DAS55 benchmark, RGCR reports improved structural self-consistency, sequence recovery, and foldable diversity. Controlled interventions investigate the contributions of joint sampling, learned utility, and recurrent feedback.

Critical strengths
  1. A well-motivated structured generation framework with useful theoretical analysis. The approach directly addresses the loss of within-step dependencies in factorized masked generation. The normalized sampling construction and the decomposition of predictive gains provide a coherent connection between the proposed method and its underlying objective.

  2. Strong empirical results supported by targeted interventions. RGCR improves structural self-consistency and sequence recovery over the evaluated baselines. Importantly, the matched-marginal experiment provides direct evidence that retaining nucleotide dependencies contributes substantially to the likelihood improvement. The component ablations further support the roles of learned selection and recurrent refinement.

Critical weaknesses
  1. The evaluation scope is relatively narrow. The primary comparison is conducted on 55 selected RNA targets. Although the selection criteria are documented, the limited test scope and the absence of clearly reported target-level uncertainty for the main comparisons make it difficult to assess robustness across diverse RNA structural settings.

  2. The relationship between predicted utilities and optimal grouping is not fully validated empirically. The theoretical results characterize true expected gains, whereas generation relies on learned approximations and greedy block selection. The current ablations support their practical usefulness, but direct analysis of gain calibration and the quality of the resulting partitions would strengthen the central methodological claim.

Questions for authors

Could the authors report paired target-level uncertainty estimates for the main structural improvements? How sensitive are the results to the DAS55 inclusion criteria or to evaluation on additional RNA structures?

How accurately do the learned coupling scores predict held-out realized or expected joint-over-singleton gains? Can the authors quantify how much performance is lost due to utility-estimation error versus greedy partition selection?

3 comments

Dear Reviewer XZRa,

Thank you for recognizing the structured-generation formulation and matched-marginal evidence.

W1 & Q1: Paired uncertainty and evaluation beyond DAS55

Paired target-level intervals support RGCR's advantage over the matched controls and the structural benefit of joint sampling:

Structural comparisonMean paired scTM improvement95% paired target-level CI
RGCR versus gRNAde after validation tuning0.15600.1383 to 0.1737
RGCR versus same-encoder AR0.11950.1057 to 0.1333
Joint versus product-of-marginals sampling, fixed checkpoint0.01090.0072 to 0.0146

The baseline comparisons use validation-tuned models trained with three seeds on the 1,138/58 sequence-group split. Additional fixed-checkpoint diagnostics, including joint versus marginals, use a separate model at training update 16,000, trained before benchmark filtering and distinct from the submitted checkpoint, with its operating point fixed. All scTM results in this response equally average RhoFold+ and AlphaFold3 with equal target weights. NLL is measured at τ=1\tau=1 in nats/nt. We average designs and runs within each target before computing paired Student-t intervals.

DAS55 selection preceded generation: common input support excludes 20 multi-chain and 23 discontinuous-connectivity targets from 98 source test groups. Selection uses neither native refolding nor generated performance.

We test transfer on 49 additional supported targets: 8 CASP15, 19 RNA-Puzzles, and 22 structsim-v2 structures. They are nonoverlapping with DAS55 and sequence-cluster-separated from development data. Target selection is fixed before generation; the three fitted models per method are reused without retuning. Temperatures for training seeds 0/1/2 remain 0.3/0.3/0.5 for RGCR and 0.1/0.5/0.1 for gRNAde.

Evaluation setTargetsgRNAde scTM, mean ± SDRGCR scTM, mean ± SD95% paired CI for RGCR minus gRNAde
DAS55, common validation-tuning protocol550.144 ± 0.0050.300 ± 0.0150.138 to 0.174
Additional RNA structures490.140 ± 0.0040.229 ± 0.0060.068 to 0.110

SD is across training-seed means. All targets, including the 720-nt target, and all designs remain in evaluation, with predictor failures scored zero. The additional-set gain is 0.0892 scTM (95% CI 0.0680 to 0.1104), supporting transfer beyond DAS55 within the common input requirements.

W2 & Q2: Accuracy and calibration of learned coupling scores

On DAS55 native-prefix contexts, we compare predicted and realized joint-over-singleton gains for executable candidates from the update-16,000 checkpoint at τ=0.1\tau=0.1.

CandidatesSpearman correlationCalibration error (nats/nt)Selected-block gain (nats/covered nt)
Pairs0.1910.0690.514
Triples0.2150.1250.445

For pairs and triples separately, define per-nucleotide predicted and realized gains for a block of size jj:

g^B=CB,τ/j,gB=(log⁡QB,τ(SB∗)−∑i∈Blog⁡qi,τ(Si∗))/j.\hat g_B=C_{B,\tau}/j,\qquad g_B=\bigl(\log Q_{B,\tau}(S_B^{\ast})-\sum_{i\in B}\log q_{i,\tau}(S_i^{\ast})\bigr)/j.

CE=∑bWb∣g^‾b−g‾b∣.\mathrm{CE}=\sum_b W_b\left|\overline{\hat g}_b-\overline g_b\right|.

Candidates are binned by predicted gain g^B\hat g_B. Each target receives equal total candidate weight before pooling; WbW_b is the bin weight and bars denote weighted bin means. The 10 fixed bins have edges (−∞,−10,−3,−1,−0.3,0,0.3,1,3,10,+∞)(-\infty,-10,-3,-1,-0.3,0,0.3,1,3,10,+\infty), are left-closed/right-open, and omit empty bins. CE is reported in nats/nt.

Correlations measure ranking within candidate order and context. Positive correlations and realized selected-block gains support useful selection information; the matched-marginal intervention separately tests the structural effect of executing these groups.

W2 & Q2: Observed-native selection headroom and greedy packing

We isolate packing error by replacing greedy selection with certified-optimal weighted set packing, holding the diagnostic checkpoint, learned scores, candidate sets, and recurrence fixed. At native contexts, both selectors receive identical predictions. NLL and the score objective use τ=1\tau=1; structural rollouts use the checkpoint's fixed operating temperature.

Packing rulePredicted-score objective relative to optimum (%)Native NLL (nats/nt)scTM
Greedy98.500.60690.4091
Certified-optimal1000.60710.4090

Across 10,560 native contexts with equal target weighting, greedy's mean objective ratio is 98.50%, a 1.50% relative gap. Optimal minus greedy is -0.0001 scTM (95% CI -0.0029 to 0.0027) and 0.00019 NLL (95% CI -0.00007 to 0.00045). The small objective gap and closely matched structural scores support greedy execution.

For selection headroom, we hold the heads fixed at τ=1\tau=1 and use certified-optimal packing for both learned-score and realized-native-gain objectives. Both partitions are evaluated on the same native assignment.

Partition comparisonMean log-score headroom (nats/nt)95% paired target-level CI
Hindsight realized-gain partition minus learned-score partition0.04130.0390 to 0.0435

The 0.0413 nats/nt measures observed-native hindsight headroom with the solver held optimal. Because the same native outcome determines selection and scoring, it includes outcome-specific variability and does not isolate expected-utility estimation error.

Thank you to the authors for the response—most of my concerns have been addressed, and I still believe my current score is appropriate and will not make further changes.

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

The manuscript proposes a masked diffusion-based method for RNA 3D inverse folding. It constructs geometry-based nucleotide pairs and triples, learns singleton and joint probability distributions, and estimates the predictive gains of joint generation over independent sampling. Experiments show improvements over some baselines.

Critical strengths

The paper presents an RNA sequence generation method combining geometric information, joint probability prediction, and recurrent refinement to explicitly model dependencies between nucleotides.

Critical weaknesses
  • Prior works such as [1-3] have studied dependency loss caused by independent sampling in discrete diffusion models, and joint generation or dependency-aware decoding. Recurrent message passing and higher-order geometric modeling have also been widely explored. I suggest a more thorough discussion of these related works to better clarify the novelty of the proposed method.

  • All training objectives in RGCR, including the supervision of its key predictive gain, are derived from the prediction of native nucleotide identities. However, the goal of RNA inverse folding is to find sequences that fold into a target structure, rather than recover a known RNA sequence. For example, most RNA 2D inverse folding methods directly optimize structural objectives, and recent 3D methods [4-5] have also explored direct structural optimization.

  • Better recovery of native sequences does not necessarily lead to better structural consistency. RGCR achieves a sequence recovery rate of ~60%, meaning that a substantial proportion of nucleotides differ from the native sequence. Without structural feedback, how can the model ensure that these differences do not disrupt the target structure?

  • The theoretical analysis also focuses mainly on sequence likelihood rather than the actual structural design objective.

  • RGCR uses the difference between joint and independent log-likelihoods on a single native sequence as the regression target for expected gain. However, each specific structure typically has only one native sequence for supervision, and these targets also change as the predictive model is trained. Can the learned gains reliably reflect the actual benefits of joint sampling? The paper also lacks a direct analysis of gain estimation accuracy and calibration.

  • Table 3-C shows that updating the block selection scores only reduces NLL from 0.599 to 0.597 when the final prediction heads are fixed. This suggests that much of the improvement may come from better probability predictions. How much does the recurrent gain feedback itself contribute?

  • RGCR learns which nucleotides should be generated jointly based on geometric information, but the positions generated at each step are determined by a predefined random schedule. Only positions assigned to the same step can be sampled jointly. So even strongly dependent positions may not be generated together. To what extent does this random schedule limit the proposed joint generation mechanism?

[1] Liu et al., Discrete Copula Diffusion. ICLR.

[2] Hersche et al., Locally Coherent Parallel Decoding in Diffusion Language Models. ICML.

[3] Ringel et al., Dependency-Guided Parallel Decoding in Discrete Diffusion Language Models.

[4] Hu et al., RIDER: 3D RNA Inverse Design with Reinforcement Learning-Guided Diffusion. ICLR.

[5] Si et al., Structure-based RNA Design by Step-wise Optimization of Latent Diffusion Model. AAAI.

Questions for authors

Please see Weaknesses.

3 comments

Dear Reviewer BAZX,

Thank you for the questions about novelty, structural design, and the contributions of predictive utility and recurrence.

W1: Novelty relative to dependency-aware generation and recurrent geometric models

RGCR connects higher-order geometric prediction and recurrent refinement (Sections 2.1 and 2.3) through predictive utility: sequence beliefs and predicted gains inform geometric messages, and the gains select joint sampling groups. Building on Section 2.2, DCD augments diffusion marginals with copula dependencies; CoDiLA uses auxiliary autoregressive local decoding; DEMASK controls parallel unmasking through predicted dependencies. RGCR selects geometry-derived pairs and triples within a fixed reveal set. Its normalized 16-way and 64-way heads supply joint distributions whose predicted gains guide both refinement and execution.

Equation 18 distinguishes predictive benefit from dependence strength:

GB=TC⁡(PB)+∑i∈BKL⁡(Pi,∣,qi)−KL⁡(PB,∣,QB).G_B = \operatorname{TC}(P_B) + \sum_{i \in B} \operatorname{KL}(P_i \\,\\|\\, q_i) - \operatorname{KL}(P_B \\,\\|\\, Q_B).

A dependent group can be unhelpful when its joint predictor is inaccurate. Submitted Table 3B supports utility-based selection: with the checkpoint and packing procedure fixed, learned gain improves scTM from 0.456 to 0.480 over TC-greedy.

W2, W3 & W4: Native supervision and the structural design objective

Native supervision trains a structure-conditioned sequence distribution; design success is evaluated by refolding complete generated sequences and comparing their structures and base pairs with the target. A non-native nucleotide need not be a design error because multiple sequences can support one backbone.

The submitted matched-marginal intervention tests coordinated sampling directly. Its scTM averages RhoFold+, AlphaFold3, and DRfold2; NLL is measured at τ=1\tau=1 in nats/nt:

Sampling kernel from the same RGCR checkpointNative NLL (nats/nt)scTMBP F1
Product of the selected heads' marginals0.6700.4480.377
Joint sampling0.5970.4800.390

At a fixed context, this intervention preserves selected groups and their one-site marginals. During rollout, each arm conditions on its own generated prefix. Improvements in scTM and BP F1 therefore connect dependence retention to structural self-consistency.

Additional fixed-checkpoint diagnostics use a separate model at training update 16,000, trained before benchmark filtering and distinct from the submitted checkpoint and the retrained fits below. With RhoFold+ and AlphaFold3 equally weighted, joint sampling scores 0.4091 versus 0.3982, a gain of 0.0109 scTM (95% CI 0.0072 to 0.0146). NLL decreases from 0.6704 to 0.6069, a reduction of 0.0635 nats/nt (95% CI 0.0578 to 0.0692).

The theory establishes normalized generation and conditional predictive-risk identities; these interventions supply the structural evidence. RIDER and SOLD pursue structural-reward optimization, complementary to RGCR's dependence-retention mechanism. The refolding results connect the generation mechanism to computational structural self-consistency.

W5: Single-native gain supervision and calibration

Each native label supplies an observation of joint-over-singleton log-score gain:

ue(Se∗;c)=log⁡Qe(Se∗∣c)−∑i∈elog⁡qi(Si∗∣c).u_e(S_e^{\ast};c) = \log Q_e(S_e^{\ast} \mid c) - \sum_{i \in e} \log q_i(S_i^{\ast} \mid c).

The shared predictor pools observations across structures, masked contexts, and candidates. With predictive functions and features fixed, Appendix A.4 gives C∗(Z)=E[ue∣Z]=E[G(V)∣Z]C^*(Z)=\mathbb{E}[u_e\mid Z]=\mathbb{E}[G(V)\mid Z], where VV identifies context, candidate, and refinement stage, and ZZ contains the predictor's features. During training, targets are recomputed from the current heads at each stage and detached in the gain loss.

We compare predictions with realized gains from the same frozen heads for executable candidates in DAS55 native-prefix contexts, using the update-16,000 checkpoint at τ=0.1\tau=0.1.

CandidatesSpearman correlationCalibration error (nats/nt)Selected-block gain (nats/covered nt)
Pairs0.1910.0690.514
Triples0.2150.1250.445

For pairs and triples separately, define per-nucleotide predicted and realized gains for a block of size jj:

g^B=CB,τ/j,gB=(log⁡QB,τ(SB∗)−∑i∈Blog⁡qi,τ(Si∗))/j.\hat g_B=C_{B,\tau}/j,\qquad g_B=\bigl(\log Q_{B,\tau}(S_B^{\ast})-\sum_{i\in B}\log q_{i,\tau}(S_i^{\ast})\bigr)/j.

CE=∑bWb∣g^‾b−g‾b∣.\mathrm{CE}=\sum_b W_b\left|\overline{\hat g}_b-\overline g_b\right|.

Candidates are binned by predicted gain g^B\hat g_B. Each target receives equal total candidate weight before pooling; WbW_b is the bin weight and bars denote weighted bin means. The 10 fixed bins have edges (−∞,−10,−3,−1,−0.3,0,0.3,1,3,10,+∞)(-\infty,-10,-3,-1,-0.3,0,0.3,1,3,10,+\infty), are left-closed/right-open, and omit empty bins. CE is reported in nats/nt.

Spearman correlations are computed within candidate order and context; selected-block gain averages realized gains of executed blocks per covered nucleotide. Positive ranking information and selected-block gains support practical utility-based selection.

W6: Contribution of recurrent gain feedback

The 0.599 to 0.597 NLL comparison isolates final routing with already-refined heads. At a fixed final partition, updating the heads reduces NLL from 0.604 to 0.597. The broader submitted controls support combined belief–utility refinement: Full achieves 0.480 scTM versus 0.431 for hidden-only recurrence with learned-gain execution and 0.438 for one-shot RGCR.

To isolate utility inputs beyond beliefs, the additional belief-only control retains three stages, belief features, prediction heads, auxiliary supervision, and learned-gain final selection. It removes utility inputs to messages and replaces utility-dependent gates with learned order-specific constants. Both fits use the 1,138/58 sequence-group split, training seed 0, equal update budgets, and validation-only operating-point selection. Full uses Torch 2.9.1/R6000 and belief-only uses Torch 2.6/A40. Their scTM equally averages RhoFold+ and AlphaFold3.

Feedback into representation refinementscTMNative NLL (nats/nt)
Beliefs only0.30050.9523
Beliefs and utilities0.28670.9548

Full minus belief-only is -0.0138 scTM (95% CI -0.0199 to -0.0078); belief-only minus Full is -0.0025 NLL (95% CI -0.0236 to 0.0187).

W7: Joint-generation opportunities under the random schedule

For a candidate of size jj, the probability of co-reveal under the random balanced schedule is

Pr⁡(e lies in one reveal set)=∑t(∣Ut∣j)(Lj).\Pr(e \text{ lies in one reveal set}) = \frac{\sum_t \binom{|U_t|}{j}}{\binom{L}{j}}.

For sufficiently long sequences and TT balanced steps, this approaches 1/T1/T for pairs and 1/T21/T^2 for triples. The diagnostic checkpoint and original schedules give the following coverage, with all-masked gains corrected to the deployed temperature:

Candidate or interaction setCo-revealed in one step (%)Executed jointly among co-revealed groups (%)
Pair candidates with positive all-masked predicted gain23.7316.33
Triple candidates with positive all-masked predicted gain5.3022.43
Target-annotated base pairs represented in the candidate graph23.8866.10

A pair is jointly executed when both positions belong to one sampled pair or triple; a triple requires one sampled triple block. Target annotations are used only for this diagnostic. The 66.10% rate is conditional on graph membership and co-reveal; 23.88% of graph-represented annotated pairs are co-revealed.

Groups split across steps can influence refinement while fully masked, and later predictions condition on generated positions. The matched-marginal structural gain demonstrates useful joint execution within these scheduling opportunities.

Reviewer_DTPfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose RGCR, a masked-diffusion language model used for fixed-backbone RNA sequence design. The structure gets encoded via a GVP encoder and then processed via recurrent updates. They then show higher scTM scores compared to gRNAde.

Critical strengths

The theory is laid out nicely and is clean and easy to follow. The ablations are well chosen and highlight that some choices such as joint sampling matter in this case. The authors also clearly highligh the quality-compute tradeoff.

Critical weaknesses

There are a few weaknesses that would strengthen the paper if addressed: [W1] It sounds like the temperature was tuned for RGCR on validation and the same was not done for the baslines, proper baseline tuning would be critical to ensure fair comparisons.

[W2] The pass rates only deduplicate based on exact sequence matches. In protein modelling often people adopt sequence-based or structure based clustering to report clustered pass rates, which seems appropriate here as well to avoid rewarding memorization.

[W3] It is not clear if the gains reported are actually statistically significant or if this is within noise since no confidence intervals etc are reported.

Questions for authors

Do the gains based on your method hold with seeds and CIs, matched temperature and budget, a clustered split, and an AR baseline on the same encoder?

2 comments

Dear Reviewer DTPf,

Thank you for recognizing the theoretical formulation, controlled interventions, and quality–compute analysis.

W1, W3 & Q1: Fair tuning, uncertainty, matched budget, and the same-encoder AR control

RGCR retains a positive paired structural advantage over gRNAde and the same-encoder AR control after matched validation tuning. All three models are retrained on the same split of 1,138 training and 58 validation sequence groups with three training seeds. Each selects temperature from {0.1, 0.3, 0.5, 1.0} on the same 26 supported validation targets, using 16 unranked designs per target and temperature, generation seed 1, and target-averaged RhoFold+ scTM. Test evaluation uses 16 unranked designs per target for each of three generation seeds. These additional scTM comparisons equally average RhoFold+ and AlphaFold3, with equal target weights and zero scores for predictor failures; they supplement the submitted three-predictor evaluation.

Both development counts use sequence groups, each represented by one conformation. Screening retains 1,138 of 4,012 training groups, with 2,800 excluded and 74 quarantined for unresolved provenance or chemical information; validation retains 58 of 98 groups, excluding 40. We screen every member against a frozen benchmark union covering DAS, structsim, RNA-Puzzles, CASP15, RNAGym, RNA3DB test, and temporal candidates. A source/PDB-chain, exact-sequence, near-duplicate-sequence, or structural match excludes the entire group. Sequence near-duplicates require identity ≥99% and target coverage ≥99% in either of the two same-orientation query/target search directions. Structural matches require benchmark-normalized TM ≥0.45, benchmark coverage ≥80%, and at least 20 aligned C4′ atoms.

The selected temperatures are:

MethodTraining seed 0Training seed 1Training seed 2
RGCR0.30.30.5
gRNAde0.10.50.1
Same-encoder AR0.50.10.3

The AR control matches the GVP encoder architecture, geometric inputs, hidden width, training data, target exposure, optimizer-update budget, and checkpoint-selection protocol, with comparable decoder capacity.

Retrained modelscTM95% paired CI for RGCR minus model
gRNAde0.144 ± 0.0050.138 to 0.174
Same-encoder AR0.180 ± 0.0220.106 to 0.133
RGCR0.300 ± 0.015Not applicable

Mean ± SD describes the three training-seed means. We average designs and repeated runs within each target, then form paired differences and Student-t 95% CIs. The CIs summarize target-level differences, while SD summarizes variation across training seeds.

For compute matching, training-seed-0 fits receive 6,600 ms per target and generation seed on the same NVIDIA A40 GPU, using temperatures 0.3, 0.1, and 0.5 for RGCR, gRNAde, and AR, respectively. The budget is fixed from validation timing and includes geometry encoding, prediction, refinement, selection, and completed-sequence transfer; refolding is separate. All completed designs are evaluated without ranking.

MethodMean completed designs per target and seedscTM at the matched generation-time budget
gRNAde39.370.150
Same-encoder AR21.760.156
RGCR81.040.288

At this budget, RGCR improves scTM over gRNAde by 0.1379 (95% CI 0.1216 to 0.1542) and over AR by 0.1314 (0.1162 to 0.1467), while producing 2.06× and 3.72× as many designs, respectively. These comparisons support both structural quality and generation throughput against the matched controls.

W2 & Q1: Clustered pass rates and sequence-cluster separation

The submitted foldable-diversity metric retains repeated sequences and all sequence pairs in its denominator, weighting structurally valid pairs by normalized Hamming distance (Appendix B, Eq. 54). Identical pairs contribute zero distance.

We additionally cluster designs at 90% identity by complete linkage, separately within each target, training run, and generation seed. A sequence medoid selected without structural scores represents each cluster. Success uses the submitted RhoFold+ scTM threshold of 0.45. Clustered pass rate is the fraction of representatives that pass; valid-cluster count measures distinct successful groups per 16 designs.

MethodClustered pass rate (%)Valid clusters per 16 designs
gRNAde0.000.00
Same-encoder AR0.140.02
RGCR18.082.89

To test training-set similarity separately, we jointly cluster development and test sequences at at least 80% identity and 80% alignment coverage of both sequences. Test clusters are excluded from training and validation, and validation clusters from training. We retain DAS55 and retrain one fit per method on the resulting 950/58 training/validation sequence-group split. Validation-selected temperatures are 0.1, 0.1, and 1.0 for RGCR, gRNAde, and AR, respectively.

MethodscTM on the sequence-cluster-disjoint split95% paired CI for RGCR minus method
gRNAde0.1490.093 to 0.114
Same-encoder AR0.1390.103 to 0.124
RGCR0.252Not applicable

RGCR retains a positive paired structural advantage after excluding close training-sequence neighbors. Together, the output-cluster and split analyses address successful-design redundancy and training–test similarity.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.