acceptodds
Under review as a conference paper at ICLR 2027

PIPER: Multi-Conformer Reasoning with Physics-Informed Regularization for RNA–Ligand Interaction Prediction

Abstract

RNA–ligand interaction prediction is an important task in RNA-targeted drug discovery, yet it remains challenging because RNAs can adopt multiple conformational states while high-resolution tertiary-structure supervision is scarce. Most existing deep learning approaches adopt a single-state formulation, compressing each RNA–ligand pair into a static representation before prediction. Collapsing multiple latent conformational states into a single pair-level representation can blur state-specific interaction signals, making it difficult to compare alternative conformational hypotheses or impose state-specific geometric and physical constraints. To address these limitations, we introduce PIPER, a physics-informed framework for explicit reasoning over a compact set of surrogate conformers without requiring atomistic structures. Given RNA sequence, secondary structure, and ligand graphs, PIPER represents each RNA–ligand pair as a compact set of surrogate conformers rather than a single static state. It performs conformer-wise competition by integrating prior-guided conformer ranking with sparse RNA-residue–ligand-atom interactions, retains the two highest-utility surrogate conformers through a sparse Top-2 posterior, and aggregates their conformer-specific scores. A source-aware screened residual regularizes the local RNA–ligand interaction regions for all generated surrogate conformers. Extensive evaluations show that PIPER outperforms strong baselines across standard benchmarks and achieves particularly strong generalization under RNA cold-start, scaffold-hopping, and dual-cold settings.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
RNA–ligand interaction prediction, multi-conformer reasoning, sparse latent-state inference, biomolecular machine learning, molecular representation learning, out-of-distribution generalization, molecular interaction modeling

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Y3dFOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses RNA–ligand interaction prediction in the structure-scarce regime. Rather than collapsing each RNA–ligand pair into a single fused representation, PIPER generates a small set of learned surrogate conformers by deforming a residue-level scaffold, ranks them using prior-guided geometric evidence and ligand-conditioned residue–atom interactions, and aggregates the top-two states for the final score. A source-aware screened residual acts as an auxiliary regularizer. Evaluations span multiple RNA–ligand benchmarks and controlled distribution-shift protocols (including RNA cold-start and scaffold-level shifts). Reported ranking metrics improve over the matched baselines, and the ablations indicate that both the sparse multi-conformer competition and the residual term contribute to the gains.

Critical strengths
  1. The problem formulation is well motivated. A single static representation can obscure state-dependent interaction signals; explicit competition among a compact set of surrogate geometric hypotheses is a reasonable inductive bias under limited tertiary-structure supervision.
  2. The empirical evaluation is broader than a single benchmark. Multiple datasets, several distribution-shift protocols, and targeted component ablations (conformer count, prior, cross-interaction, residual variants) make the comparative case more convincing.
  3. The authors state the limitations of the geometry and field clearly: the conformers are coarse surrogate geometries rather than atomistic ensembles, and the neural fields supply physics-informed regularization rather than calibrated electrostatic predictions. This framing is appropriate and should be retained.
Critical weaknesses
  1. The surrogate conformers are not shown to correspond to real conformational states. They arise from learned local deformations of a simple scaffold, and the paper provides no comparison with experimentally resolved RNA structures or ligand-specific binding modes/contact patterns. Consequently, the central “multi-conformer reasoning” claim is stronger than the evidence supports; the method is more accurately described as inference over a set of latent geometric hypotheses.
  2. The physics-informed component is less persuasive than the ranking results. While the screened residual has a plausible form, its contribution depends on several design choices (source assignment, probe sampling, screening) that are only partially controlled. Existing ablations (generic L2, homogeneous PDE residual, shuffled sources) show degradation, yet they do not fully isolate specificity from a generic smoothness or capacity effect of an auxiliary field regularizer.
  3. Evidence for consistent superiority is uneven. Some gains are modest relative to typical run-to-run variation, and the paper does not report clear target-level or paired statistical analysis. Claims that the method consistently outperforms strong baselines should therefore be stated more cautiously.
Questions for authors
  1. Can the learned surrogate conformers be compared with any resolved RNA structures or ligand-specific contact patterns to demonstrate that they capture meaningful interaction-relevant states?
  2. What additional control would more cleanly separate the contribution of the source-aware screened residual from a generic smoothness or capacity effect of an auxiliary field?
  3. Which performance gains remain significant across random seeds and individual RNA targets, and how sensitive are the results to the number of retained conformers (Top-1 / Top-2 / All-3)?
4 comments
AuthorsOpenReview ↗

Response to Reviewer Y3dF (Part 1 / 3)

We sincerely appreciate the reviewer’s positive and insightful feedback, particularly the recognition of PIPER’s well-motivated formulation, multi-conformer reasoning, and comprehensive evaluation. Your encouragement further motivates our research.

We have carefully considered your constructive comments and questions, and strengthened the supporting evidence and methodological clarity. We provide detailed point-by-point responses below.


W1 and Q1. Structural Evidence for Multi-Conformer Reasoning

We thank the reviewer for suggesting validation against experimentally resolved RNA structures. Following your suggestion, we conducted an external structural assessment on HARIBOSS to examine whether PIPER's surrogate conformers capture interaction-relevant information. As defined in Section 3.1, these conformers represent latent geometric hypotheses rather than recovered molecular structures.

External structural assessment. We evaluated 114 RNA–ligand complexes with experimentally resolved contact information and 123 complexes for geometric assessment. Experimental structures were used exclusively for post hoc evaluation, without influencing model training or conformer selection. We compared PIPER with a matched no-screened-residual ablation, examining whether posterior preferences align with structure-derived RNA–ligand contacts.

For this diagnostic analysis, the selected conformer is defined by the highest posterior probability; PIPER's original Top-2 aggregation for interaction prediction remains unchanged, as defined in Eq. (13).

Table R3.1. External structural assessment on HARIBOSS.

Structure-derived metricPIPERNo residual
Best-contact candidate selection (%)87.740.4
Posterior–contact ranking (Spearman)+0.50−0.10
Selected-candidate contact AUPR0.29960.2998
Selected pocket RMSD (Å)15.6515.65

Contact-based analyses use 114 complexes; geometric evaluation uses 123 complexes.

Evidence of contact-consistent reasoning. PIPER selects the candidate most consistent with experimental contacts in 87.7% of complexes, compared with 40.4% for the no-residual variant. Among discordant paired outcomes, 64 favor PIPER versus 10 favor the ablation (two-sided exact sign test, p = 8.96 × 10⁻¹¹). Moreover, the Spearman correlation between posterior preference and contact-derived candidate ranking increases from −0.10 to +0.50. These results provide external structural evidence that PIPER's posterior preferences better align with experimentally observed interaction contacts, with screened-residual regularization contributing to this alignment.

Structural interpretation. This improvement concerns the relative selection of interaction-relevant hypotheses rather than atomistic reconstruction. The selected candidates show comparable absolute contact AUPR (0.2996 vs. 0.2998) and pocket RMSD (15.65 Å for both models). Thus, the findings support contact-consistent hypothesis ranking within PIPER's multi-conformer framework, consistent with its design of evaluating competing surrogate geometries through interaction evidence rather than recovering physical conformational states.

The complete HARIBOSS assessment and its interpretation are included in Appendix B.17 (Table 27) of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer Y3dF (Part 2 / 3)

W2 and Q2. Isolating Screening from Generic Regularization

We thank the reviewer for suggesting this targeted control. Following your advice, we conducted a source-preserving, no-screening experiment to isolate the contribution of screening from generic regularization and additional model capacity.

Matched screening control. As defined in Eq. (14), PIPER employs the residual

Resk(p)=−∇2ϕk(p)+κ2ϕk(p)−ρk(p)\mathrm{Res}_k(p) = -\nabla^2 \phi_k(p) + \kappa^2 \phi_k(p) - \rho_k(p)

where ρk\rho_k denotes the surrogate source terms and κ\kappa controls screening. Setting κ=0\kappa = 0 removes only the screening term while retaining the source-aware Poisson constraint. Unlike the homogeneous-PDE ablation in Table 2, this control preserves the source assignments, Laplacian operator, and auxiliary-field capacity.

We independently retrained three variants on the SMRTnet fixed split, holding the data partition, architecture, source construction, probe configuration, regularization weight, optimizer, hyperparameters, and checkpoint-selection procedure fixed.

Table R3.2. Matched screening controls on SMRTnet.

Screening configurationAUCAUPR
PIPER (screened residual, κ = 0.15)0.86670.7466
Source-preserving, no screening (κ = 0)0.85840.7364
Stronger screening (κ = 0.30)0.85820.7349

Contribution of screening. Removing screening alone reduces AUC/AUPR by 0.0083/0.0102, despite identical model capacity, source assignments, and regularization settings. Stronger screening (κ = 0.30) also performs below the default, indicating that increased screening strength does not necessarily improve predictive performance.

Distinction from generic regularization. The existing generic L2, homogeneous-PDE, and shuffled-source controls (Table 2; Appendix B.12) examine different aspects of field regularization, differential constraints, and source–geometry alignment. The new κ = 0 control specifically isolates screening while preserving the source terms. Together, these comparisons provide stronger evidence that PIPER's predictive benefit is associated with the source-aware screened formulation, rather than additional capacity or generic regularization alone.

These results support a positive contribution from screening under matched conditions. The screening-specific matched control and its implementation are now reported in Appendix B.12 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer Y3dF (Part 3 / 3)

W3 and Q3. Multi-Run Robustness and Conformer-Support Sensitivity

We thank the reviewer for emphasizing multi-run evidence and posterior-support sensitivity. We address both concerns through expanded OOD comparisons, paired RNA-clustered analysis, and three-run Top-1/Top-2/All-3 evaluation.

1. Robustness against strong baselines. To assess whether PIPER's performance advantages persist across independent runs, we extended the evaluation to three independent training runs for PIPER and all six external baselines across four OOD protocols, using matched splits and reporting mean ± SD. PIPER achieves the highest mean AUC/AUPR under RNA cold-start (0.8659/0.7480), scaffold-hopping (0.7327/0.5563), and dual cold-start (0.6827/0.4658).

We further conducted paired bootstrap analysis (5,000 resamples) to quantify the reliability of these improvements. Under RNA cold-start, PIPER outperforms the strongest baseline, RNAsmol, by +0.1769 AUC (95% CI [+0.1562, +0.1974]) and +0.2436 AUPR (95% CI [+0.2089, +0.2748]). Under scaffold-hopping, it surpasses SDNE by +0.0499 AUC (95% CI [+0.0240, +0.0754]) and +0.0494 AUPR (95% CI [+0.0174, +0.0839]). All four paired confidence intervals exclude zero, providing evidence that these improvements are not attributable solely to test-set sampling variability.

These evaluations support PIPER's generalization advantages beyond single-run comparisons. Complete three-run results and paired statistical analyses are provided in Appendices B.6–B.8 of the revised manuscript.

2. Matched ablations and RNA-level uncertainty. We compared PIPER against independently trained single-conformer and no-screened-residual variants across all four OOD protocols. To account for dependencies among pairs sharing the same RNA, we performed paired RNA-clustered bootstrapping (2,000 resamples per run), resampling complete RNA groups.

Table R3.3. Three-run OOD differences and RNA-clustered uncertainty.

OOD protocolPIPER vs.ΔAUCAUC CI excl. 0ΔAUPRAUPR CI excl. 0
Ligand coldSingle conformer+0.00851/3+0.00061/3
No residual+0.02242/3+0.02581/3
RNA coldSingle conformer−0.00421/3−0.00871/3
No residual+0.02633/3+0.02732/3
Scaffold-hoppingSingle conformer+0.00111/3+0.00121/3
No residual+0.02023/3+0.00040/3
Dual coldSingle conformer+0.03012/3+0.02301/3
No residual+0.01661/3+0.00700/3

Δ denotes the three-run mean PIPER-minus-variant difference. CI counts indicate how many run-specific paired RNA-clustered 95% CIs exclude zero, regardless of direction; they do not establish cross-run significance.

The stronger evidence concerns screened-residual regularization. Removing it reduces mean AUC/AUPR across all four protocols. Under RNA cold-start, PIPER improves AUC/AUPR by +0.0263/+0.0273, with clustered CIs excluding zero in 3/3 and 2/3 runs. Under scaffold-hopping, the AUC improvement is +0.0202, with CIs excluding zero in all three runs. In contrast, multi-conformer gains are protocol-dependent: they are largest under dual cold-start, modest under ligand cold-start and scaffold-hopping, and slightly negative under RNA cold-start. Full paired and RNA-clustered analyses are included in Appendix B.8.

3. Posterior-support sensitivity. We extended the Top-1/Top-2/All-3 comparison to three independent runs. All variants generate three surrogate conformers and differ only in the number retained for posterior aggregation. Top-1 is distinct from the single-conformer ablation in Table 2, which generates only one conformer and disables conformer conditioning.

Table R3.4. Posterior-support sensitivity on SMRTnet (mean ± SD).

Retained supportAUCAUPR
Top-10.8708 ± 0.01330.7625 ± 0.0227
Top-2 (PIPER)0.8781 ± 0.00310.7753 ± 0.0052
All-30.8726 ± 0.00370.7668 ± 0.0064

Top-2 achieves the highest mean AUC/AUPR and lowest observed SD, improving by +0.0073/+0.0128 over Top-1 and +0.0055/+0.0085 over All-3. These results favor sparse Top-2 aggregation among the tested configurations, without establishing cross-run statistical superiority (Appendix B.16).

Overall, the expanded analyses provide stronger paired evidence for PIPER's advantages under RNA cold-start and scaffold-hopping, while identifying protocols where gains are modest or uncertain. The ablations further support screened-residual regularization, and the three-run support comparison favors Top-2 aggregation. We have incorporated these results into the revised manuscript and clarified the protocol-dependent scope of our conclusions.


We hope our responses have addressed your concerns and welcome any further feedback. We would be happy to incorporate these discussions into the camera-ready version.

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer Y3dF,

Thank you again for your thoughtful review and constructive questions. As the discussion period approaches its close, we would like to briefly follow up on our responses and the revised manuscript.

Following your suggestions, we have completed additional three-run OOD evaluations with paired uncertainty analyses (Section 4.3; Appendices B.6–B.8), screening and geometry controls (Appendices B.12–B.13), and an external HARIBOSS structural assessment (Section 4.5; Appendices B.16–B.17). We have also refined the Abstract and Conclusion to clarify the protocol-dependent contributions of multi-conformer reasoning and the interpretation of surrogate coordinates. All corresponding revisions are highlighted in blue.

We would greatly appreciate it if you could take another look and let us know whether these updates address your concerns. We would be happy to clarify any remaining questions before the discussion closes.

Best regards, The Authors

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

This paper introduces PIPER, a framework for predicting RNA–small-molecule interactions from RNA sequences, secondary structures, and ligand graphs without requiring resolved three-dimensional structures. The model generates three surrogate RNA conformers, evaluates them using geometric priors and ligand-specific interaction information, and combines predictions from the two highest-scoring conformers. Experiments show strong predictive performance and generalization, with ablations supporting the benefits of multi-conformer reasoning and physics-informed regularization.

Critical strengths

The paper presents a novel approach to RNA–ligand interaction prediction by comparing three surrogate RNA conformers without requiring resolved 3D structures. Evaluation across three benchmarks and multiple distribution shifts shows particularly strong RNA cold-start performance, while ablation studies support the effectiveness of the key components.

Critical weaknesses

The surrogate conformers lack structural validation, leaving their relationship to real RNA conformations uncertain and limiting biological interpretation.

Questions for authors
  1. How sensitive is performance to the choice of surrogate geometry?
  2. Where experimental structures are available, could you assess whether the surrogate conformers capture meaningful RNA structural features?
3 comments
AuthorsOpenReview ↗

Response to Reviewer dxFC (Part 1 / 2)

We sincerely appreciate the reviewer’s positive and insightful feedback, particularly the recognition of PIPER’s novel multi-conformer reasoning framework and comprehensive experimental evaluation. Your encouragement further motivates us to advance this research.

We have carefully considered your constructive feedback and strengthened the supporting evidence and methodological clarity. We provide point-by-point responses below.


Q1. Sensitivity to Surrogate Geometry

We thank the reviewer for this important question. Our controlled experiments show limited predictive sensitivity to the tested surrogate coordinate constructions and scales, while preserving the benefit of screened-residual regularization.

Controlled geometry comparison. Following your suggestion, we independently retrained matched variants on the SMRTnet fixed split, evaluating all models on the same 2,469 held-out examples. The data partition, training seed, architecture, optimizer, hyperparameters, and checkpoint-selection procedure were held fixed.

We compared the original RNA coordinates [i/NR,sin⁡(i/2),cos⁡(i/3)]\left[i/N_R,\sin(i/2),\cos(i/3)\right] and planar ligand circle defined in Eqs. (2) and (3) with an alternative construction using [i/NR,sin⁡(i/3),cos⁡(i/5)]\left[i/N_R,\sin(i/3),\cos(i/5)\right] and a 0.25sin⁡(2θ)0.25\sin(2\theta) vertical displacement for ligand atoms. Both constructions were evaluated with and without screened-residual regularization, retaining neural-field features in the no-residual controls. We additionally tested coordinate scales of 0.5×0.5\times and 2×2\times. These experiments use newly trained models.

Table R2.1. Surrogate geometry and scale sensitivity on SMRTnet (AUC/AUPR).

SettingAUCAUPR
Original chart (PIPER)0.86670.7466
Alternative chart (PIPER)0.86360.7467
Original chart (No residual)0.84710.7261
Alternative chart (No residual)0.84230.7184
Original chart (0.5× scale)0.86590.7426
Original chart (2× scale)0.86170.7500

Geometry and scale sensitivity. Across the tested constructions and scales, PIPER achieves AUCs of 0.8617–0.8667 and AUPRs of 0.7426–0.7500. The original-minus-alternative differences are only +0.0031 AUC (95% CI [−0.0083, +0.0138]) and −0.0001 AUPR (95% CI [−0.0239, +0.0229]). These results indicate modest performance variation within the tested geometric perturbations.

Preserved regularization benefit. Importantly, screened-residual regularization improves AUC/AUPR by +0.0196/+0.0205 under the original chart and +0.0213/+0.0283 under the alternative. Both paired AUC improvements have 95% CIs excluding zero ([+0.0066, +0.0323] and [+0.0079, +0.0342], respectively). The alternative chart also yields an AUPR improvement with 95% CI [+0.0012, +0.0546]. These findings show that the regularization benefit is not confined to the original sinusoidal frequencies or planar ligand coordinates.

Conclusion. Consistent with Sections 3.1 and 3.5, PIPER treats surrogate conformers as latent geometric hypotheses rather than experimentally resolved molecular structures. The additional experiments support predictive robustness across the tested geometric variations without implying physical invariance. The geometry and scale analyses, confidence intervals, and implementation details are included in Appendix B.13 (Tables 21–22) of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer dxFC (Part 2 / 2)

W1 and Q2. Structural Evidence from Resolved RNA–Ligand Complexes

We thank the reviewer for suggesting validation against experimentally resolved RNA structures. Following your suggestion, we conducted an external structural assessment on HARIBOSS to examine whether PIPER's surrogate conformers capture interaction-relevant information. As clarified in Section 3.1, these conformers represent latent geometric hypotheses rather than atomistically resolved structures.

External structural assessment. We evaluated 114 RNA–ligand complexes with experimentally resolved contact information and 123 complexes for geometric assessment. Experimental structures were used exclusively for post hoc evaluation, without influencing training or conformer selection. We compared PIPER with a matched no-screened-residual variant, examining whether posterior preferences align with structure-derived RNA–ligand contacts.

For this diagnostic analysis, the selected candidate is defined by the highest posterior probability; PIPER's original Top-2 aggregation for interaction prediction remains unchanged, as defined in Eq. (13).

Table R2.2. External structural assessment on HARIBOSS.

Structure-derived metricPIPERNo residual
Best-contact candidate selection (%)87.740.4
Posterior–contact ranking (Spearman)+0.50−0.10
Selected-candidate contact AUPR0.29960.2998
Selected pocket RMSD (Å)15.6515.65

Contact-based analyses use 114 complexes; geometric evaluation uses 123 complexes.

Contact-consistent conformer selection. PIPER selects the candidate most consistent with experimental contacts in 87.7% of complexes, compared with 40.4% for the no-residual variant. Among discordant paired outcomes, 64 favor PIPER versus 10 favor the ablation (two-sided exact sign test, p = 8.96 × 10⁻¹¹). Moreover, the Spearman correlation between posterior preference and contact-derived candidate ranking increases from −0.10 to +0.50. These findings provide external structural evidence that PIPER's posterior preferences better align with experimentally observed interaction contacts, with screened-residual regularization contributing to this alignment.

Structural interpretation. The improvement concerns the relative selection of contact-consistent hypotheses rather than atomistic reconstruction. The selected candidates show comparable absolute contact AUPR (0.2996 vs. 0.2998) and pocket RMSD (15.65 Å for both models), without corresponding improvements in structural accuracy. Thus, these results support interaction-relevant hypothesis ranking within PIPER's multi-conformer framework, consistent with its intended role in reasoning over competing surrogate geometries rather than recovering physical conformational states.

The complete HARIBOSS assessment and its structural interpretation are included in Appendix B.17 (Table 27) of the revised manuscript.


We hope the response has addressed your concerns. If you have any additional concerns or comments that we may have missed in our responses, we would be most grateful for any further feedback from you to help us further enhance our work. We are more than happy to include all these discussions in the camera-ready version of this work.

Reviewer_dxFCOpenReview ↗

Response

Thank you for the detailed responses. I have read the additional geometry-sensitivity experiments and the HARIBOSS assessment. I appreciate the clarification that the structural evidence supports contact-consistent candidate ranking rather than reconstruction of physical RNA conformations. I have no further questions and will maintain my original score.

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

This paper proposes PIPER for RNA–ligand interaction prediction when high-resolution tertiary structures are scarce. Given an RNA sequence, secondary structure, and a ligand molecular graph, PIPER constructs a deterministic residue-level surrogate scaffold and generates three surrogate conformers. It combines a geometry-guided ranking prior with ligand-conditioned residue–atom interaction evidence, retains the two highest-utility conformers through a sparse Top-2 posterior, and aggregates their conformer-specific energy-style scores. A source-aware screened PDE residual provides an additional conformer-specific regularizer in the surrogate coordinate space. Experiments cover SMRTnet, Robin, and Biosensor, together with four SMRTnet-derived distribution shifts: ligand cold-start, RNA cold-start, scaffold-hopping, and dual cold-start. The paper also includes component ablations and mechanistic analyses. The results are promising, especially under RNA cold-start and scaffold-hopping. My main reservations concern the physical interpretation of the surrogate-space regularizer and the reproducibility and uncertainty of the OOD results.

Critical strengths
  1. The core modeling hypothesis is clear and testable. Retaining a small set of competing surrogate conformers could preserve state-specific interaction evidence better than selecting one conformer or aggregating all candidates. The method separates conformer generation, prior-guided ranking, ligand-conditioned interaction modeling, sparse Top-2 selection, and score aggregation. The independently trained Top-1/Top-2/All-3 comparison (Table 19), together with ablations of the prior, cross-graph interaction, and screened-residual loss (Table 2), provides useful evidence about the contributions of these components. The L2, homogeneous-PDE, and shuffled-source controls also make the regularization claim more specific than a comparison against no regularization alone.

  2. The evaluation covers three RNA–ligand benchmarks, four distribution-shift protocols, and several baselines. The authors report validation-set threshold selection and Holm-adjusted paired DeLong tests for AUC on the main benchmarks (Appendices B.6–B.7). They also explicitly state that the coordinates and neural fields are surrogate constructs, rather than atomistic structures or calibrated electrostatic predictions. This qualification helps define the method’s intended scope and should remain central to interpreting its results.

Critical weaknesses
  1. The physical interpretation of the screened-residual term is not yet persuasive. The RNA scaffold uses an engineered sequence-index chart, x_i = [i/N_R, sin(i/2), cos(i/3)] (Eq. 2), while ligand atoms are assigned a deterministic circular reference chart (Eq. 3). The source term uses unit signed surrogate sources rather than atom-specific partial charges (Appendix B.5). The authors appropriately acknowledge that this is not a calibrated atomistic electrostatic model, and the ablations support the regularizer’s predictive usefulness. Still, the evidence more directly supports this particular engineered regularizer than a constraint that captures meaningful molecular physics. The existing controls distinguish the complete residual from several alternatives, but do not establish how its benefit depends on the arbitrary chart construction. Sensitivity to alternative coordinate constructions and coordinate scaling would help clarify which physical interpretation, if any, the results support. This is a concern about attribution and framing, rather than evidence that the regularizer is ineffective.

  2. The strongest distribution-shift claims are difficult to audit from the submitted package, and their robustness is not quantified clearly enough. The package documents a runnable fixed-split SMRTnet protocol and supplies the primary split files. It also contains generic exact-RNA and scaffold grouping helpers (train_piper.py, lines 591–654), so split-construction code is not entirely absent. However, I could not identify the exact sample-ID manifests or the implementation of the four clustering-based OOD protocols described in Appendix B.6. Those generic helpers do not reproduce the stated ligand-similarity clustering, RNA 5-mer clustering, merged scaffold groups, or joint dual-cold assignment. The archive also lacks prediction files, checkpoints, or per-run logs linking these evaluations to the reported numbers. Tables 6–7 give OOD point estimates without run-to-run variability or uncertainty intervals. This is particularly relevant for dual cold-start, whose test partition has 255 pairs (Table 5). Its AUPR gain over the matched single-conformer variant is 0.0116 (0.4864 vs. 0.4748), while the no-screened-residual-loss variant has higher AUC than PIPER (0.7030 vs. 0.6931, Table 7). These results do not invalidate the method, but they leave the strength, stability, and component attribution of the OOD conclusions less certain than the paper’s wording suggests.

Questions for authors
  1. Can you provide the exact sample-ID manifests and construction scripts for the four SMRTnet-derived OOD partitions, together with an audit of RNA-sequence, ligand-identity, and scaffold overlap across train, validation, and test? Appendix B.6 outlines the splitting rules and reports aggregate statistics, but the included artifacts do not directly reproduce the reported protocols. Please clarify how the generic grouping helpers in the code relate to the clustering and exclusion rules used for Tables 6–7.

  2. Were multiple random seeds used for Tables 6–7? If so, please report means with uncertainty estimates, such as standard deviations across runs and appropriate confidence intervals, and paired comparisons on the same held-out examples where applicable. If the results are from a single run, please state that limitation explicitly. This matters particularly for the small dual-cold-start test set and the modest AUPR gaps between PIPER and its matched ablations.

  3. Since the RNA and ligand coordinates are surrogate charts rather than molecular 3D coordinates, how sensitive is the screened-residual benefit to the choice and scale of these charts? Please clarify which parts of the residual have a physical rationale and provide a controlled comparison using alternative coordinate constructions. The L2, homogeneous-PDE, and shuffled-source controls are informative, but alone do not show that the gain arises from physically meaningful geometry rather than task-specific regularization.

6 comments
AuthorsOpenReview ↗

Response to Reviewer uTmr (Part 1 / 3)

We sincerely appreciate the reviewer’s positive and insightful feedback, particularly their recognition of PIPER’s clear and testable modeling hypothesis, sparse multi-conformer reasoning, and comprehensive evaluation across RNA–ligand benchmarks and distribution shifts. Your encouragement further motivates us to advance this research.

We have carefully addressed your concerns through additional analyses. We hope these efforts help strengthen your confidence in our findings and respectfully invite you to reconsider your rating. Below, we provide detailed responses to each comment.


W2 and Q1. Exact OOD Manifests and Reproducibility Audit

We thank the reviewer for emphasizing the importance of reproducible OOD evaluation. Following your suggestion, we have made an anonymous reproducibility bundle available at https://osf.io/m8yrx/overview?view_only=1b772149559d449badb88e447df594b4. The bundle includes exact sample-ID manifests, partition construction scripts, overlap audits, archived predictions, evaluation metrics, training logs, available checkpoints, and replay scripts for all five evaluation protocols, together with the SHA-256 checksum of the original SMRTnet input file. Importantly, the replay script independently recomputes AUC and AUPR from archived predictions, reproducing all 40 model–protocol result entries in Tables 7–8 of the revised manuscript.

Partition construction and overlap audit. The build_splits and build_group_splits functions in train_piper.py handle group-based partition assignment, while protocol_utils.py implements the protocol-specific rules, including Morgan-fingerprint ligand clustering, RNA-sequence grouping, Bemis–Murcko scaffold grouping, and joint ligand/RNA-group disjointness with mixed-assignment exclusion for dual cold-start.

We further audited exact RNA-sequence, ligand-identity, and molecular-scaffold overlap between training and test partitions.

Table R1.1. Exact-entity overlap between training and test partitions.

ProtocolRNA overlapLigand overlapScaffold overlap
Fixed split1,794278216
Ligand cold-start1,780010
RNA cold-start0262202
Scaffold-hopping1,71700
Dual cold-start007

All sample-ID assignments are disjoint. Complete train–validation, train–test, and validation–test overlap audits are included in the anonymous bundle.

The audit confirms zero overlap for the entity groups explicitly held out by each OOD protocol. Scaffold overlap under ligand and dual cold-start is permitted by their definitions, while the row-level fixed split allows entity overlap by design.

Conclusion. The anonymous reproducibility bundle enables independent verification of the exact sample assignments, disjointness constraints, and reported evaluation metrics. Partition statistics and overlap checks are now documented in Appendix B.6 (Tables 4–6); the code will be made publicly available upon acceptance.

AuthorsOpenReview ↗

Response to Reviewer uTmr (Part 2 / 3)

W2 and Q2. Multi-Run OOD Robustness and Paired Uncertainty

We thank the reviewer for highlighting the importance of uncertainty quantification in OOD evaluation. Following your suggestion, we extended the originally single-run results in Tables 6–7 to three independent training runs for PIPER, all six external baselines, and both matched ablations across four OOD protocols.

All comparisons use fixed partitions and identical held-out examples. We report mean ± SD across runs and compute paired 95% confidence intervals (CIs) using 5,000 class-stratified bootstrap resamples. Identical sampled indices are applied across models and runs before averaging paired differences. The CIs quantify test-set sampling uncertainty conditional on the trained models, complementing the across-run SD.

Table R1.2. OOD comparison with the top three baselines by mean AUPR (mean ± SD).

ProtocolModelAUCAUPR
Ligand coldPIPER0.7395 ± 0.02480.5878 ± 0.0262
DeepRSMA0.7519 ± 0.04000.5895 ± 0.0364
SDNE0.6828 ± 0.03050.5540 ± 0.0213
DeepConv-DTI0.6885 ± 0.03520.5319 ± 0.0424
RNA coldPIPER0.8659 ± 0.00470.7480 ± 0.0123
RNAsmol0.6890 ± 0.01000.5044 ± 0.0066
DeepConv-DTI0.5704 ± 0.07930.3686 ± 0.0591
BINDTI0.5425 ± 0.01260.3492 ± 0.0140
Scaffold-hoppingPIPER0.7327 ± 0.00400.5563 ± 0.0023
SDNE0.6829 ± 0.05840.5069 ± 0.0620
DeepConv-DTI0.6485 ± 0.04050.4690 ± 0.0581
BINDTI0.5940 ± 0.03200.4363 ± 0.0735
Dual coldPIPER0.6827 ± 0.01260.4658 ± 0.0191
DeepConv-DTI0.6374 ± 0.00310.4168 ± 0.0173
SMRTnet0.5531 ± 0.04340.3927 ± 0.1043
BINDTI0.5930 ± 0.04460.3857 ± 0.0372

All six external baselines were evaluated; the complete results are reported in Appendix B.6 (Table 7).

Table R1.3. OOD comparison with matched ablations (mean ± SD).

ProtocolVariantAUCAUPR
Ligand coldPIPER0.7395 ± 0.02480.5878 ± 0.0262
Single conformer0.7311 ± 0.01550.5872 ± 0.0171
No residual0.7171 ± 0.01860.5620 ± 0.0163
RNA coldPIPER0.8659 ± 0.00470.7480 ± 0.0123
Single conformer0.8702 ± 0.00080.7567 ± 0.0088
No residual0.8397 ± 0.00960.7207 ± 0.0112
Scaffold-hoppingPIPER0.7327 ± 0.00400.5563 ± 0.0023
Single conformer0.7316 ± 0.00580.5551 ± 0.0146
No residual0.7126 ± 0.01360.5559 ± 0.0148
Dual coldPIPER0.6827 ± 0.01260.4658 ± 0.0191
Single conformer0.6525 ± 0.06320.4428 ± 0.0587
No residual0.6661 ± 0.04530.4588 ± 0.0424

External-baseline robustness. PIPER achieves the highest three-run mean AUC and AUPR among all six external baselines under RNA cold-start, scaffold-hopping, and dual cold-start. The strongest advantage occurs under RNA cold-start, outperforming RNAsmol by +0.1769 AUC (95% CI [+0.1562, +0.1974]) and +0.2436 AUPR (95% CI [+0.2089, +0.2748]). Under scaffold-hopping, PIPER also improves AUPR over the strongest baseline by +0.0494 (95% CI [+0.0174, +0.0839]). These results provide strong paired evidence for PIPER's advantages under RNA and scaffold distribution shifts.

Matched-ablation robustness. Removing screened-residual regularization reduces mean AUC and AUPR across all four OOD protocols. PIPER improves AUPR by +0.0258 under ligand cold-start and +0.0273 under RNA cold-start, with paired 95% CIs excluding zero. Multi-conformer gains are more protocol-dependent: PIPER outperforms the single-conformer variant under three protocols, with the largest gains under dual cold-start (+0.0301 AUC / +0.0230 AUPR), but performs slightly worse under RNA cold-start. These results highlight the more consistent benefits of screened-residual regularization and the protocol-dependent contribution of multi-conformer reasoning.

Conclusion. The additional three-run evaluations and paired uncertainty analyses strengthen the evidence for PIPER's OOD generalization, particularly under RNA cold-start and scaffold-hopping, while clarifying the contributions of its key components. Complete three-run results, paired CIs, and experimental details are now included in Appendices B.6–B.8 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer uTmr (Part 3 / 3)

W1 and Q3. Surrogate Geometry and Screened-Residual Sensitivity

We thank the reviewer for raising this important question about the physical interpretation and geometric dependence of the screened residual.

Physical rationale. As clarified in Sections 3.1 and 3.5, PIPER represents conformers as "latent residue-level geometric hypotheses" rather than atomistic RNA structures. The screened residual (Eq. (14)) provides physics-informed regularization through a screened Poisson operator combining a Laplacian, screening term, and geometry-dependent sources. As detailed in Appendix B.4, these sources use surrogate positions and unit signed values rather than physical partial charges. Thus, the physical motivation lies in the differential constraint, while the coordinates and source assignments serve as computational inductive biases.

Controlled geometry comparison. Following your suggestion, we independently retrained four matched variants on the SMRTnet fixed split, keeping the data partition, training seed, architecture, optimizer, hyperparameters, and checkpoint-selection procedure fixed.

We compared the original RNA coordinates [i/NR,sin⁡(i/2),cos⁡(i/3)]\left[i/N_R,\sin(i/2),\cos(i/3)\right] and planar ligand circle defined in Eqs. (2) and (3) with an alternative construction using [i/NR,sin⁡(i/3),cos⁡(i/5)]\left[i/N_R,\sin(i/3),\cos(i/5)\right] and a 0.25sin⁡(2θ)0.25\sin(2\theta) vertical displacement for ligand atoms. Both constructions were evaluated with and without screened-residual regularization, retaining neural-field features in the no-residual controls. We additionally tested coordinate scales of 0.5×0.5\times and 2×2\times. These experiments use newly trained models.

Table R1.4. Sensitivity to surrogate chart construction (AUC/AUPR).

Coordinate chartScreened residualNo residualΔAUCΔAUPR
Original0.8667 / 0.74660.8471 / 0.7261+0.0196+0.0205
Alternative0.8636 / 0.74670.8423 / 0.7184+0.0213+0.0283

Preserved regularization benefit. Importantly, screened-residual regularization improves AUC/AUPR by +0.0196/+0.0205 under the original chart and +0.0213/+0.0283 under the alternative. Both paired AUC improvements have 95% CIs excluding zero ([+0.0066, +0.0323] and [+0.0079, +0.0342], respectively). The alternative chart also yields an AUPR improvement with 95% CI [+0.0012, +0.0546]. These findings show that the regularization benefit is not confined to the original sinusoidal frequencies or planar ligand coordinates.

Chart and scale sensitivity. The original-minus-alternative differences are only +0.0031 AUC (95% CI [−0.0083, +0.0138]) and −0.0001 AUPR (95% CI [−0.0239, +0.0229]). Coordinate rescaling by 0.5× and 2× yields AUC/AUPR values of 0.8659/0.7426 and 0.8617/0.7500, respectively, compared with 0.8667/0.7466 at the original scale. These results indicate limited predictive sensitivity within the tested variations, without implying physical scale invariance.

Conclusion. Together with the existing L2, homogeneous-PDE, and shuffled-source controls (Table 2; Appendix B.12), these experiments provide further evidence for the predictive contribution of structured, source-aware regularization beyond the original chart. The complete sensitivity analyses are now included in Appendix B.13 (Tables 21–22), with their interpretation clarified in Sections 3.2 and 3.5.


We are happy to see from your comments that you have an overall positive view of our work, and we respectfully ask if you'd be comfortable to adjust your ratings, based on all the new experiments and justification to your questions. Thank you once again for your time and effort. We look forward to incorporating these insightful discussions into the final camera-ready version.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all three reviewers for their thoughtful and constructive feedback. We appreciate their recognition of PIPER's motivation, multi-conformer reasoning framework, and evaluation across RNA–ligand benchmarks and distribution shifts.

We have carefully addressed the reviewers' specific concerns through targeted additional experiments, statistical analyses, and methodological clarifications. The revised manuscript retains PIPER's original framework and core contributions, with updates focused on strengthening experimental evidence and clarifying the interpretation of our findings.

Core Contributions of Our Work

  1. Sparse Multi-Conformer Reasoning: PIPER explicitly evaluates competing surrogate conformational hypotheses through geometry-guided ranking and ligand-conditioned interaction evidence, retaining the two highest-utility candidates for prediction.

  2. Physics-Informed Regularization: A source-aware screened residual provides conformer-specific regularization in surrogate geometric space, complementing multi-conformer interaction reasoning without requiring resolved RNA tertiary structures.

  3. Evaluation Across Distribution Shifts: PIPER is evaluated on three RNA–ligand benchmarks and four controlled OOD protocols, with targeted ablations and mechanistic analyses examining its predictive behavior.

Updates of Experimental Results During Rebuttal

  • Multi-Run OOD Evaluation (R1, R3): Extended PIPER, all six external baselines, and matched ablations to three independent runs across four OOD protocols. Added mean ± SD, paired bootstrap confidence intervals, and RNA-clustered uncertainty analyses (Section 4.3; Appendices B.6–B.8).

  • Screening and Geometry Controls (R1, R2, R3): Added a source-preserving no-screening control to isolate the contribution of screening, together with alternative surrogate coordinate constructions and scale-sensitivity experiments (Appendices B.12–B.13).

  • Structural and Posterior-Support Assessment (R2, R3): Added an external HARIBOSS assessment using experimentally resolved RNA–ligand contacts, and extended the Top-1/Top-2/All-3 comparison to three independent runs (Section 4.5; Appendices B.16–B.17).

Updates of Methodological Clarifications During Rebuttal

  • Physical and Structural Interpretation (R1, R2, R3): Clarified surrogate conformers as latent geometric hypotheses, surrogate coordinates as computational inductive biases, and screened-residual regularization as physics-informed rather than calibrated electrostatics (Sections 3.2, 3.5, and 5; Appendices B.13 and B.17).

  • Scope of Empirical Conclusions (R1, R3): Clarified the consistent mean benefits of screened-residual regularization, protocol-dependent multi-conformer gains, and stronger paired evidence under RNA cold-start and scaffold-hopping (Abstract; Sections 4.3 and 5; Appendix B.8).


These targeted revisions address the reviewers' concerns while preserving PIPER's original methodological scope. All changes are highlighted in blue in the revised manuscript, with detailed experimental evidence and explanations provided in our individual responses.

We thank the reviewers again for their valuable feedback and welcome any further questions or suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.