acceptodds
Under review as a conference paper at ICLR 2027

Monte Carlo Tree Hallucination for All-Atom Biomolecular Sequence–Structure Co-Design

Abstract

Biomolecular design underpins applications from molecular recognition to therapeutics and synthetic biology, yet de novo interaction design remains challenging—especially for DNA/RNA, underexplored non-protein modalities with scarce, heterogeneous complex data and sharper geometric and chemical constraints. We introduce MCTH (Monte Carlo Tree Hallucination), an inference-only framework that casts all-atom sequence–structure co-design as uncertainty-aware planning over hallucinated states from pretrained folding and inverse-folding models, with optional biophysical control within the same decision loop. MCTH treats these models as frozen black-box operators and uses Monte Carlo Tree Search to allocate a fixed inference budget across competing design trajectories, incorporating model confidence and uncertainty, as well as cross-expert consensus/disagreement when multiple predictors are available. Across protein–RNA, protein–DNA, protein–protein, and protein–ligand design, matched-budget experiments show that adaptive search improves over simpler sampling and cycling strategies, while held-out AlphaFold3 and Chai-1 evaluations demonstrate transfer beyond the search-time oracle. MCTH provides a shared planning layer across modalities while allowing task-specific folding, inverse-folding, and biophysical modules, requiring no fine-tuning or backpropagation through component models.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Biomolecular design, Sequence–structure co-design, Monte Carlo tree search, Inference-time planning, Molecular hallucination

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_6YTFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MCTH uses Monte Carlo Tree Search to allocate folding evaluations across biomolecular sequence–structure design trajectories. Each expansion proposes sequences through inverse folding, predicts their structures, and evaluates the resulting complexes. The CU-PUCT selection rule combines action values with consensus, uncertainty, and optional biophysical signals. The experiments cover RNA and DNA aptamers, protein binders, and ligand-conditioned protein design, with controlled search comparisons, held-out structure predictors, and Rosetta interface evaluation.

Critical strengths
  1. Retaining alternative sequence–structure trajectories helps MCTH use a fixed folding budget more effectively. Appendix E.3 compares search strategies under the same initialization, proposal models, scoring, and folding oracle. Across the controlled subsets of three RNA, three DNA, and two protein targets, mean best search-time ipTM increases from 0.784 with greedy cycling to 0.913 for RNA, from 0.828 to 0.955 for DNA, and from 0.928 to 0.938 for proteins (Tables 15–17). These results support the value of revisiting earlier hypotheses and refining different branches instead of committing evaluations to a single trajectory.

  2. The reported gains transfer to predictors outside the search loop. Under AlphaFold3 evaluation, MCTH achieves higher mean ipTM than ODesign for both RNA and DNA aptamers (Tables 4b and 6b). Adding AlphaFold2 during protein-binder search increases mean best Chai-1 ipTM from 0.18 to 0.57 across eleven targets, taking the best score among three retained designs per target (Table 8). Chai-1 is excluded from proposal generation, search, and archive ranking (Appendix E.6). On six shared protein-binder targets, MCTH designs generated with the physical term disabled also have area-normalized Rosetta interface energies comparable to RFdiffusion (Appendix E.7 and Table 25).

Critical weaknesses
  1. The paper does not demonstrate that CU-PUCT’s consensus, uncertainty, and physical terms improve upon a conventional tree policy using the same models and rewards. Equation 4 combines these factors in the exploration bonus, but setting the exploration constant to zero removes the entire bonus. This comparison establishes the value of exploration without showing why the proposed selection factors are needed. The dual-expert experiment also changes both the evaluation objective and the consensus/disagreement signals. Its Chai-1 improvement could therefore result from optimizing the combined Boltz-2/AlphaFold2 objective, without requiring the proposed selection rule. No direct comparison establishes the benefit of physical guidance either. The results support adaptive search, but do not establish the additional benefit of the proposed selection and guidance mechanisms.

  2. The reported final-design yield is measured on outputs selected from pooled search runs, but the aggregate folding budget used to produce these outputs is not specified. Appendix E.6 pools candidates from available dual-expert runs, removes duplicate sequences, and retains the highest-ranked candidates using the recorded search reward. RFdiffusion contributes ten generated outputs per target. The authors acknowledge that equal output counts do not imply equal compute (Table 2). The 37.3% versus 20.0% Chai-1 pass fractions establish a higher pass rate for the reported MCTH archive, but do not show whether this advantage persists under equal total generation and selection budgets. Both methods obtain at least one pass on eight of eleven targets.

Questions for authors
  1. Can the authors compare CU-PUCT with standard PUCT using the same proposals, evaluation objective, and folding-call budget? In the dual-expert setting, the control should retain both Boltz-2 and AlphaFold2 in the reward so that the comparison tests the selection rule rather than the choice of evaluation models. Separately, which physical operators were enabled in the reported DNA experiments? Comparing the existing procedure with these operators disabled would help determine whether physical guidance improves held-out scores or physical feasibility beyond confidence-based search.

  2. How many runs and total folding calls contributed to each reward-ranked archive in Appendix E.6? Does the final-design yield advantage persist when the total generation budget and the opportunity to rank and retain candidates are matched? A comparison using a common output-selection procedure, without access to Chai-1 scores, would help establish whether MCTH produces a higher fraction of qualifying designs under comparable computational resources.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 6YTF

Q1(a): “compare CU-PUCT with standard PUCT using the same proposals, evaluation objective, and folding-call budget”

We agree that setting the exploration constant to zero removes the entire bonus and does not isolate CU-PUCT from a standard tree policy. We ran the direct comparison on the 22 expanded nucleic-acid tasks at three seeds, with identical initialization, proposal models, scoring, and folding-call budget (Tables R1–R2, Global Response): CU-PUCT is above PUCT by +0.0145 (p = 0.035) and above UCT by +0.0188 (p = 0.0071).

Q1(b): separating the selection rule from the evaluation objective

“The dual-expert experiment also changes both the evaluation objective and the consensus/disagreement signals. Its Chai-1 improvement could therefore result from optimizing the combined objective, without requiring the proposed selection rule.”

We ran the 2×2 control. The second expert (OpenFold3) is either absent, used only in the reward, used only in the consensus/disagreement term, or both. The reward only arm keeps the combined evaluation objective but uses a standard selection rule.

Table R4. Dual-expert ablation, 11 DNA targets. Following the Table 8 protocol: one design is selected from each of three seeds (11, 23, 47); each predictor column reports the highest ipTM among those three, and the mean averages the 11 target-level maxima. The maximizing design may differ across predictors. All arms use 100 inverse-folding and 100 folding calls per target per seed. AF3 and Chai-1 are excluded from proposal generation, search, and ranking.

SettingAF3Δ vs fullChai-1Δ vs full
Dual: consensus + reward0.6564—0.5372—
Dual: reward only0.6391−0.01730.5148−0.0224
Dual: consensus only0.6127−0.04360.5055−0.0317
Single expert0.5891−0.06730.5081−0.0291
ODesign0.4936−0.16270.4431−0.0941

The full configuration is best under both predictors. The gain over a single expert (+0.0673 AF3, +0.0291 Chai-1) comes from both components: the combined evaluation objective contributes +0.0500 / +0.0067, and the consensus term a further +0.0173 / +0.0224. Per target, the full rule beats reward only on 6 of 11 targets under AF3 and 7 of 11 under Chai-1; at this sample size the result is directional rather than statistically resolved.

The consensus term helps when the second expert is also in the reward, but not on its own: consensus only stays close to the single-expert baseline (+0.0236 AF3, −0.0026 Chai-1). This matches the intended mechanism, since disagreement between experts carries information only once both actually score the candidates.

Why the second expert enters late. Both experts run only after the first 60 folding calls; earlier expansions use a single expert, so the total budget matches the single-expert arm at 100 calls. The switch point is fixed across all targets and seeds, and was not tuned per target or against any held-out predictor. Two considerations motivate it. First, marginal returns per folding call are steepest early (Table 3B: +0.110 from 1 to 5 calls, +0.034 from 50 to 100), so spending two calls per node at that stage would halve the number of distinct candidates explored where exploration pays most. Second, early candidates are largely poor and both experts tend to score them low together, so the disagreement signal carries little information until candidates reach a range where the experts can differ. The ablation above is consistent with the second point.

Per-target results for all four configurations under both predictors are available and we will add them to the appendix.

Q1(c): “which physical operators were enabled in the reported DNA experiments?”

The submitted DNA experiments enabled only the ΔG_self reward penalty at ω = 0.1; sequence editing was not enabled. We ran the comparison the reviewer asks for — the same procedure with the penalty disabled, and with sequence editing enabled — in Tables R5–R6 under 739X W2.

On whether physical guidance improves held-out scores or physical feasibility beyond confidence-based search: physical feasibility, not confidence. Disabling the penalty gives the highest mean best ipTM (0.6887 at ω = 0), while enabling it with sequence editing gives the best post-repacking RosettaDNA score (+181.4 vs +323.5) and the fewest clashes (7 vs 17). Since the penalty adds a term to the optimized objective, we report the trade-off rather than only the configuration with the highest confidence score.

Q2: “How many runs and total folding calls contributed to each archive? Does the yield advantage persist under matched budgets and a common selection procedure?”

We should have specified this protocol in the paper. Both MCTH and RFdiffusion were evaluated under a budget of 100 folding calls per target, with 10 final designs retained from each pipeline. MCTH ranks candidates with its in-loop joint reward; RFdiffusion uses its native output-selection procedure. Neither method uses Chai-1 for candidate selection; Chai-1 is applied afterward as a common held-out evaluator.

Under this protocol MCTH achieves 41/110 passing designs (37.3%) against 22/110 (20.0%) for RFdiffusion, a 1.86× higher pass rate. Both methods achieve at least one passing design on 8 of 11 targets.

The folding budget and the output count were therefore already matched, with each method using its own native selection mechanism. We will state this in the revision, since the submission reported the pass rates without the budget behind them.

On the dual-expert setting for protein binders: not covered here, for the compute reasons in the Global Response.

Reviewer_6YTFOpenReview ↗

Official Comment by Reviewer 6YTF

Thank you for the additional results. I have acknowledged them. The DNA/OpenFold3 ablation is useful, but I still cannot determine how much of the original protein-binder improvement comes from CU-PUCT rather than the combined Boltz-2/AlphaFold2 reward. Could you check the existing logs and report how many runs went into each pooled archive, the total folding calls, and whether the 100-call budget includes all those runs? Please also describe the pass-rate result as a comparison of the two complete pipelines, as their output-selection procedures differ.

Reviewer_TRrZOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Monte Carlo Tree Hallucination (MCTH), an inference-time tree search framework for biomolecular sequence–structure co-design. It combines uncertainty, cross-expert consensus, and optional biophysical constraints without retraining the underlying models. Experiments show promising results across several design tasks.

Critical strengths
  1. A general inference-time planning framework for biomolecular co-design. The paper formulates sequence–structure co-design as a search problem and uses MCTS to allocate expensive folding evaluations adaptively.
  2. The authors compare MCTH against independent sampling, greedy cycling, and exploration-disabled search under controlled folding-call budgets. They also evaluate candidates using predictors that are not used during search. The study therefore distinguish the benefits of adaptive search from simple repeated evaluation and predictor-specific optimization.
Critical weaknesses
  1. I'm worring about the reproducibility. Limited benchmark coverage and insufficient evidence of statistical robustness. The evaluation uses relatively small target sets. This makes the improvements less reliable.
  2. The contributions of individual planning and biophysical components remain insufficiently established. MCTH combines several mechanisms, including tree-based exploration, uncertainty bonuses, consensus priors, and optional physical penalties. Although the paper includes useful search ablations, it acknowledges that not every scoring term is isolated experimentally. It is therefore difficult to determine which components are essential, whether the full CU-PUCT formulation consistently outperforms simpler search policies, and whether the additional complexity is justified by measurable gains. Furthermore, matching the number of folding evaluations does not necessarily match total computational cost, since candidate generation and other pipeline operations may differ. This limits the strength of the efficiency and methodological-necessity claims.
Questions for authors
  1. The matched-budget ablations against greedy cycling, best-of-N, and exploration-off baselines are useful and support the benefit of search-guided compute allocation. However, the individual contributions of CU-PUCT's components remain unclear. Could you add ablations that separately remove the consensus prior, inverse-folding uncertainty, and folding uncertainty? The optional biophysical penalty could also be ablated in tasks where it is enabled.
  2. Given the relatively small benchmark sets and limited seed coverage, can the authors report per-target variability and uncertainty estimates across independent runs? In particular, does MCTH consistently outperform the strongest baseline on held-out targets, or are the aggregate gains driven by a small number of favorable cases?
  3. The paper reports good results on independent predictors such as AlphaFold3 and Chai-1, but the supplementary analysis also indicates that agreement between Boltz-2 and AlphaFold3 at the individual-design level varies across modalities. How sensitive are the conclusions to the choice of held-out predictor, and can the authors clarify the extent to which the reported improvements reflect better interface design rather than differences in predictor-specific scoring preferences?
4 comments
AuthorsOpenReview ↗

Response to Reviewer TRrZ

Q1: “the individual contributions of CU-PUCT’s components remain unclear”

We agree the original search ablations do not isolate each term, and we ran all four removals.

The two uncertainty terms are complementary (Table R1, Global Response). Removing inverse-folding uncertainty costs −0.0109 (CI [−0.0284, +0.0051]) and removing folding uncertainty costs −0.0116 (CI [−0.0265, +0.0023]); removing both while keeping exploration costs −0.0226 (CI [−0.0399, −0.0053]). The two signals carry overlapping but not identical information, and together account for most of the margin over a standard policy.

The consensus prior was confounded with the evaluation objective in the submission, so we ran a 2×2 control on 11 DNA targets under two held-out predictors (Table R4, under 6YTF Q1b). The full configuration is best under both predictors, and the gain over a single expert comes from both components: the combined objective contributes +0.0500 under AF3 and +0.0067 under Chai-1, the consensus term a further +0.0173 and +0.0224. The consensus term helps when the second expert is also in the reward, but not on its own.

The biophysical penalty changes what the search optimizes, so removing it is not a like-for-like comparison. On ipTM it costs −0.0155 at ω = 0.1; on physical plausibility it raises the in-window fraction from 27.6% to 37.3% and improves the post-repacking RosettaDNA score from +323.5 to +282.0, and with sequence editing to +181.4 with clashes down from 17 to 7 (Tables R5–R6, under 739X W2).

Q2: “small benchmark and limited seed coverage”

We now report 22 nucleic-acid tasks with three seed-conditioned trees per task, with per-target mean ± SD below and target-level bootstrap intervals in Tables R1–R2. On whether the gains come from a few favourable cases: CU-PUCT is above ODesign on 20 of 22 targets, and the two exceptions (8VX4_dna35, 8X0N_rna12) are cases where both methods already exceed 0.78.

TargetCU-PUCTODesign
DNA (11)
B7H3_dna300.539 ± 0.1020.381 ± 0.048
CD3D_dna300.552 ± 0.1340.253 ± 0.064
FGFR2_dna300.515 ± 0.0590.414 ± 0.012
PDL1_dna300.462 ± 0.0640.402 ± 0.017
TNFR1_dna300.526 ± 0.0140.443 ± 0.031
7UV7_dna210.902 ± 0.0300.853 ± 0.024
7UXY_dna320.899 ± 0.0400.873 ± 0.049
8C58_dna140.744 ± 0.0310.723 ± 0.002
8SLN_dna290.710 ± 0.0300.605 ± 0.024
8U0P_dna110.831 ± 0.0530.760 ± 0.023
8VX4_dna350.918 ± 0.0170.920 ± 0.007
RNA (11)
B7H3_rna300.513 ± 0.1290.426 ± 0.100
CD3D_rna300.336 ± 0.0510.293 ± 0.065
FGFR2_rna300.512 ± 0.0680.407 ± 0.017
PDL1_rna300.590 ± 0.1110.496 ± 0.060
TNFR1_rna300.514 ± 0.0440.375 ± 0.009
8SWB_rna200.898 ± 0.0210.868 ± 0.011
8SWC_rna200.896 ± 0.0150.863 ± 0.034
8X0N_rna120.782 ± 0.0140.840 ± 0.022
9NY9_rna190.755 ± 0.0480.682 ± 0.059
9RVP_rna340.466 ± 0.0230.398 ± 0.082
9XZR_rna710.613 ± 0.1090.417 ± 0.070

The de novo targets carry larger seed variance (SD up to 0.134) than the NA-12 targets, and they are also where the margin over ODesign is largest. We will state this in the revision.

Q3: “Sensitivity to the held-out predictor”

We evaluated the full 22-task benchmark under two held-out predictors (Table R7, under 739X W4). CU-PUCT is ahead of ODesign under both AF3 (+0.0422) and Chai-1 (+0.0551) and on both modality subsets, with no reversal.

The direction is consistent across two predictors from different developers, which argues against a preference specific to one scorer. Absolute values still differ between predictors (0.4737 vs 0.4240 for CU-PUCT), so we do not read any particular ipTM level as meaningful on its own. We also report an evaluation independent of folding-model confidence — post-repacking RosettaDNA geometry and clash counts (Table R6, under 739X W2) — where the sequence-editing configuration improves while scoring lower on ipTM.

AuthorsOpenReview ↗

A Gentle Follow-up

Dear Reviewer TRrZ,

Thank you again for your thoughtful and detailed review. We have addressed each concern you raised and included additional experiments and clarifications in our responses. With less than 24 hours left in the discussion period, we would welcome any further feedback and would be grateful if you would reconsider your initial rating in light of our responses.

Best regards, The Authors

Thanks for your additional experiments and explainations. The new experiments addressed my questions on contributions of uncertainty terms and consensus prior, and the expanded target set addressed my concern about stability of the gain. I therefore raised my score to 3.

One last question, could you clarify the statement of "The biophysical penalty changes what the search optimizes, so removing it is not a like-for-like comparison."?

(Two additional suggestions: 1) navigate to "Table xxx, under xxx)" caused some trouble to read and understand, it would be better if you could paste the same table to every reviewer, 2) could you reference to the sections that you updated in the manuscript, and highlight the changes?)

Dear Reviewer TRrZ,

Thank you for reconsidering your score and for asking us to clarify. Setting ω = 0 is a valid ablation, and our original wording was imprecise. We meant that evaluating the ablation on ipTM alone captures the change in model confidence but leaves out whether the penalty increases the fraction of nodes within the specified ΔG_self window.

At ω = 0, the node reward depends only on folding-model confidence. At ω > 0, the reward also accounts for deviation from the ΔG_self window (−10 to −2 kcal/mol). A node with slightly lower confidence can therefore receive a higher reward if its predicted folding free energy is closer to the window.

The first two rows isolate the effect of the penalty at the same folding budget, with sequence editing disabled. Across 11 DNA targets, enabling the penalty reduced mean best ipTM by 0.0155 and increased the fraction of in-window nodes by 9.7 percentage points. The post-repacking RosettaDNA score and clash count also improved. Neither metric was used in the search reward.

ConfigurationNodes in window ↑Mean best ipTM ↑Mean RosettaDNA score after repacking ↓Clash pairs <2 Å after repacking ↓
ω = 027.6%0.6887+323.517
ω = 0.137.3%0.6732+282.014
ω = 0.1 with sequence editing39.9%0.6554+181.47

The third row shows the additional effect of enabling sequence editing at fixed ω = 0.1. Because each configuration used one seed per target, we interpret these differences as an observed trade-off.

Thank you also for both presentation suggestions. We will consider to include the relevant tables directly in each response and highlight the changes in the revised manuscript. We will clarify the trade-off in §3.3, report the full ω sweep in Appendix C.2, and describe sequence editing in Appendix C.7. Appendix E.3 will cover the UCT/PUCT comparison and component ablations, and Appendix E.6 will cover the dual-expert control and matched archive protocol. We will also highlight the expanded 22-task benchmark in Appendix A and the held-out predictor evaluations in Appendices E.2 and E.5.

Best regards, The Authors

Reviewer_739XOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces Monte Carlo Tree Hallucination (MCTH), an inference-only framework for biomolecular sequence–structure co-design. MCTH coordinates frozen folding and inverse-folding models through Monte Carlo Tree Search, allocating a limited structure-prediction budget across competing design trajectories. Its CU-PUCT selection rule incorporates model confidence, uncertainty, cross-expert consensus, and optional biophysical penalties. Experiments cover protein–RNA, protein–DNA, protein–protein, and protein–ligand design, with additional evaluations using predictors excluded from the corresponding search loops.

The problem is practically relevant, and the controlled experiments provide encouraging evidence for adaptive search. However, the current evaluation does not sufficiently establish the contribution of the proposed selection mechanism beyond standard tree search. Limited benchmark coverage and unmatched runtime in the final-design comparison also weaken the empirical conclusions.

Critical strengths
  • The problem is well-motivated and the proposed method is generally reasonable. Structure-prediction calls are expensive, making their adaptive allocation an important practical question. Retaining alternative sequence–structure trajectories offers a reasonable mechanism for recovering from unproductive refinement paths. The method combines existing folding and inverse-folding models without fine-tuning or differentiation through their parameters. Demonstrating the same planning interface across several molecular settings is a useful aspect of the work.

  • The effectiveness of the proposed method is verified via extensive controlled comparisons. The comparisons against independent sampling, greedy cycling, and exploration-off search hold initialization, proposal models, scoring, and folding-call budgets fixed. These experiments provide evidence that adaptive allocation can improve optimization beyond repeated sampling alone.

  • The held-out AF3 and Chai-1 evaluations, together with Rosetta interface analysis, strengthen the empirical study. I appreciate the reporting of cross-predictor disagreement and the acknowledgment that computational confidence does not establish experimental binding.

Critical weaknesses
  • The specific algorithmic contribution is insufficiently isolated. The main search comparisons establish the usefulness of exploration and branching, but do not show whether CU-PUCT improves over standard UCT or PUCT. Setting the exploration constant to zero removes the entire exploration term, simultaneously disabling several proposed mechanisms. Consequently, the results cannot distinguish the benefit of ordinary tree search from that of uncertainty, consensus, or physical modulation. Comparisons with standard MCTS and separate component-removal experiments are central to assessing the methodological contribution.

  • The evidence for biophysics–ML co-hallucination is limited. The DNA free-energy penalty modulates the exploration bonus, while its independent effect on final design quality is not established. With the reported weight of 0.1 and a penalty in [0,1], its multiplicative factor ranges only from approximately 0.905 to 1. This makes a direct sensitivity analysis important for understanding its practical effect. The manuscript should also identify which experiments enable sequence refinement and structural guidance, and quantify their individual contributions. The Rosetta results support competitive physical plausibility, but do not demonstrate superior interface energetics.

  • Benchmark coverage and statistical support are insufficient for broad generalization claims. Five of the seven RNA tasks involve sequences of only 4–8 nucleotides. Their representativeness of functional aptamer design therefore requires justification. The DNA and ligand benchmarks also contain multiple chain-level tasks derived from relatively few PDB complexes. Moreover, the main comparisons largely use one random seed, while the controlled protein search comparison covers only two targets. This is consequential given the small mean advantage over Complexa in protein-binder design (0.864 versus 0.853). More independent targets and repeated runs are needed to establish the robustness of these gains.

  • Cross-predictor robustness and compute efficiency remain only partially established. The held-out evaluations are valuable, but the reported results also reveal substantial predictor dependence. In Table 8, single-expert MCTH achieves mean ipTM values of 0.69 under Boltz-2 and 0.18 under Chai-1; Appendix E.5 reports weak within-target agreement for RNA designs. These findings limit the interpretation of search-time improvements. In addition, the equal-output comparison with RFdiffusion does not match total compute: MCTH selects an archive from search candidates, including candidates pooled across available runs, whereas RFdiffusion contributes ten generated outputs. The reported pass-rate improvement therefore supports final-archive quality under the stated protocols, rather than a matched-compute efficiency advantage.

Questions for authors

Please refer to the weaknesses pointed above.

A convincing standard-MCTS comparison, clearer evidence for the individual mechanisms, and repeated held-out evaluations under matched computational budgets may improve the overall quality.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 739X

We thank the reviewer for a review that identified the central weakness precisely, including the arithmetic on the physical-penalty weight, which we had not stated ourselves.

W1: “The specific algorithmic contribution is insufficiently isolated.”

We agree this was the paper’s main gap, and we ran both parts the reviewer asks for.

For the comparison against standard MCTS, see Tables R1–R2 in the Global Response: CU-PUCT is above PUCT by +0.0145 (p = 0.035) and above UCT by +0.0188 (p = 0.0071) on 22 targets × 3 seeds, with everything except the selection rule held fixed.

For the separate component removals: the two uncertainty terms are in Table R1 and are complementary (each alone ≈ −0.011 with an interval including zero; both together −0.0226, CI [−0.0399, −0.0053]). The consensus prior is isolated in Table R4 (under 6YTF Q1b), and the physical penalty is swept in Tables R5–R6 below.

W2: “The evidence for biophysics–ML co-hallucination is limited.”

The reviewer’s arithmetic was correct and pointed to the right experiment: with weight 0.1 and a penalty in [0,1], the multiplicative factor spans only ≈0.905–1.0, so an on/off test would have told us little. We report a sweep over ω, and separately the sequence-editing operator.

The ΔG window. The penalty uses a window ΔG_self ∈ [−10, −2] kcal/mol with soft bounds (k_l = k_u = 0.75), where ΔG_self is the designed strand’s intramolecular MFE computed from sequence alone (seqfold, SantaLucia 1998, 37 °C). A node is in-window if its ΔG_self lies inside this interval. The bounds follow the calibration in §C.2 and the design prior in §3.3.

Why ipTM is not the only axis. Model confidence measures the folding model’s self-estimate of its own prediction; it does not track thermodynamic cost. We therefore report both.

Table R5. Physical penalty coefficient sweep. All 11 DNA targets, one seed; search-time oracle. ω = 0 is the base CU-PUCT configuration of Table R3 (DNA); the difference (0.6887 vs 0.6908) is because this sweep uses one seed rather than three.

ωMean best ipTM ↑In-window nodes ↑Targets whose top-ipTM node is in-window ↑
00.688727.34/11
0.10.673236.95/11
0.30.675644.46/11
1.00.669344.47/11

This sweep uses a single seed per target. Each arm still receives the full budget, but five configurations × 11 targets at three seeds did not fit the discussion period, so we read it as a trend.

Raising ω costs 0.019 ipTM in total while raising the in-window node count from 27.3 to 44.4 and moving the best-scoring node itself into the window on 7/11 rather than 4/11 targets.

Sequence editing. The reviewer asked which experiments enable sequence refinement versus structural guidance. The submitted DNA experiments used the reward penalty at ω = 0.1 and did not enable sequence editing. The two operators act at different levels: ω changes which nodes the tree prefers to visit, sequence editing changes which nodes the tree can generate.

The ordinary tree proposes four child sequences per expansion with NA-MPNN. The seq-edit tree keeps the first three and, when eligible, replaces the fourth with a single-base edit of the parent sequence, so both trees evaluate four children per expansion on the same folding budget. An edit is attempted only when the parent’s ΔG_self falls outside the window; editable positions must lie at least 8 Å from every receptor heavy atom in the parent’s predicted structure; among single-base substitutions that reduce the window deviation, the best one under seqfold is taken. If none qualifies, the fourth child stays as proposed.

Table R6. Sequence-editing ablation. Same 11 DNA targets and seed as Table R5; search-time oracle. ODesign is scored over its 100 generated candidates and our arms over 99 expanded nodes, so in-window figures are fractions.

MethodIn-window ↑Mean best ipTM ↑Mean RosettaDNA score after repacking ↓<2 Å clash pairs after repacking ↓
ODesign13.9%0.5980+418.517
ω = 027.6%0.6887+323.517
ω = 0.137.3%0.6732+282.014
seq edit, ω = 0.139.9%0.6554+181.47

Sequence editing sits further along the same trade-off than the penalty alone: the largest in-window fraction and the best post-repacking geometry, at the lowest ipTM. Because edits are restricted to positions ≥8 Å from every receptor heavy atom, the improvement in interface score and clash count is not obtained by optimizing interface residues directly.

On the reviewer’s point that our Rosetta results showed comparable, not superior, interface energetics: these scores let us compare designs against each other under one protocol, and they show the physical term reducing interface energy and clashes (+418.5 → +181.4; 17 → 7). We do not interpret these values as ΔG estimates and do not claim they establish binding.

W3: “Benchmark coverage and statistical support are insufficient for broad generalization claims”

We agree the original 4–8 nt RNA tasks do not support broad claims about aptamer design. The expanded benchmark is described in the Global Response: 22 tasks × 3 seeds, chains of 11–71 nt, with per-target SD, bootstrap intervals, and paired tests.

On the thin protein-binder margin over Complexa (0.864 vs 0.853): we agree, and we did not add protein-binder targets in this period. We will present that comparison as inconclusive rather than as support for the method. The ligand benchmark is likewise unchanged; see the Global Response for the compute reason.

W4: “Cross-predictor robustness and compute efficiency remain only partially established”

The reviewer is right that the Boltz-2 0.69 / Chai-1 0.18 gap in Table 8 limits what our search-time numbers establish. We evaluated the full expanded benchmark under two predictors excluded from proposal generation, search, and ranking.

Table R7. Held-out predictor evaluation. 22 tasks × 3 seeds; mean ipTM across seeds.

PredictorMethodDNA (11)RNA (11)All 22
AF3ODesign0.41310.44990.4315
AF3CU-PUCT0.45540.49190.4737
Chai-1ODesign0.36980.36800.3689
Chai-1CU-PUCT0.41910.42900.4240
The gap largely closes here: 0.4737 under AF3 against 0.4240 under Chai-1, a difference of 0.05 rather than 0.51. We attribute the original gap to the small number of very short RNA targets rather than to the method, consistent with W3.

The ordering is also stable: CU-PUCT is ahead under both predictors and both modalities, with no reversal on any of the four subsets. Absolute values sit well below the search-time oracle, as expected once the scoring model leaves the loop.

On compute matching for the final-design comparison, see our response to Reviewer 6YTF Q2.

AuthorsOpenReview ↗

A Gentle Follow-up

Dear Reviewer 739X,

Thank you again for your thoughtful and detailed review. We have addressed each concern you raised and included additional experiments and clarifications in our responses. With less than 24 hours left in the discussion period, we would welcome any further feedback and would be grateful if you would reconsider your initial rating in light of our responses.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.