acceptodds
Under review as a conference paper at ICLR 2027

RiboSBDD and RiboSBDDBench: Structure Based Small Molecule Generation for RNA

Abstract

Structure based drug design (SBDD) for RNA is limited by scarce RNA ligand structures and flexible binding pockets. We introduce **RiboSBDD**, an equivariant diffusion framework for RNA conditioned 3D ligand generation. Despite different backbones, interfaces between ligands and proteins or RNA share local packing principles. A unified all atom representation lets us transfer these geometric priors through protein pretraining and RNA specific fine tuning. To learn local pocket adjustment without receptor trajectories, we restrict motion to nucleobase rotations about glycosidic bonds with a fixed backbone. Synthetic rotations provide analytical inverse targets. This couples base correction with ligand denoising through a differentiable pocket transformation. To retain protein derived geometric guidance after RNA specialization, an adaptive sampler combines coordinate proposals from the frozen protein source and RNA target using local equal density composition and potential guided particle resampling. We leave atom types to the target. We also establish **RiboSBDDBench**, with 8,006 curated RNA pocket ligand instances, grouped structural splits, and evaluation across multiple metrics. On 713 test pockets, RiboSBDD improves contact recovery and steric compatibility over TargetDiff after RNA fine tuning. Together, RiboSBDD and RiboSBDDBench provide a generative framework that combines transfer with local flexibility and a standardized evaluation foundation for RNA targeted small molecule design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI4S, Molecule Design, Drug Discovery, RNA

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_oPU9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents RiboSBDD, an SE(3)-equivariant diffusion framework for RNA-conditioned 3D small-molecule generation, addressing data scarcity and RNA pocket flexibility. The framework combines protein-to-RNA geometric transfer, local nucleobase correction, and adaptive source–target sampling to improve ligand generation. The authors also introduce RiboSBDDBench, a curated dataset containing 8,006 RNA–ligand pocket instances with structured data splits and evaluation protocols. Experiments on 713 test pockets demonstrate improved contact recovery, reduced steric clashes, and better AnnapuRNA interaction scores compared with RNA-fine-tuned TargetDiff.

Critical strengths

S1. The paper addresses RNA-targeted structure-based drug design, an important yet relatively underexplored area. Tackling data scarcity and RNA pocket flexibility is meaningful for extending molecular generation methods beyond protein targets.

S2. The framework combines protein-to-RNA knowledge transfer with local nucleobase correction, providing a reasonable approach to leveraging protein-derived geometric priors while accounting for RNA-specific structural characteristics.

S3. RiboSBDDBench provides 8,006 curated RNA–ligand instances with grouped structural splits and multiple evaluation metrics. This dataset could serve as a useful resource for developing and evaluating future RNA-targeted molecular generation methods.

S4. The experiments demonstrate improvements in RNA–ligand spatial compatibility, particularly in contact recovery, steric clash reduction, and AnnapuRNA scores. Comparisons across four baseline architectures and the module ablations provide useful evidence supporting the proposed framework.

Critical weaknesses

W1. Local Base Correction is trained using synthetic nucleobase rotations rather than observed ligand-induced conformational changes. Although it improves interaction metrics, direct evidence that the model learns biologically meaningful RNA flexibility is still lacking. The improvements may also partly arise from additional geometric supervision.

W2. The current ablation removes coordinate composition and particle resampling simultaneously, making their individual contributions unclear. Comparisons with simpler strategies, such as fixed-weight averaging or adaptive mixing without resampling, would better demonstrate the necessity of the proposed approach.

W3. While RiboSBDDBench provides a useful dataset, further analysis of RNA target diversity, ligand interaction quality, and neighboring protein contributions would strengthen its reliability. Evaluation on an independent dataset or downstream task would also help demonstrate its practical utility.

W4. The specific methodological improvements of RiboSBDD over TargetDiff and their individual contributions could be more clearly demonstrated. While RiboSBDD achieves better spatial compatibility, it shows reduced synthetic accessibility and molecular diversity. More controlled comparisons and detailed analyses would help clarify the advantages of RiboSBDD over TargetDiff, the sources of its performance gains, and the trade-offs involved.

Questions for authors

Q1. Could the authors provide more direct evidence that Local Base Correction captures meaningful RNA conformational changes, such as angular prediction errors or comparisons with apo–holo structures? Could they also distinguish its contribution from that of the additional geometric losses?

Q2. How does the proposed adaptive sampling method compare with simpler alternatives, such as fixed-weight coordinate averaging? Could the contributions of coordinate composition and particle resampling be evaluated separately?

Q3. Could the authors provide further analyses of RNA target diversity, ligand interaction quality, and neighboring protein effects in RiboSBDDBench? Would evaluation on an independent dataset or downstream task be feasible?

Q4. Could the authors clarify which specific methodological improvements distinguish RiboSBDD from TargetDiff and provide stronger evidence of their individual contributions? Additionally, how do these modifications explain the improvements in spatial compatibility alongside the reductions in synthetic accessibility and molecular diversity?

6 comments
AuthorsOpenReview ↗

Response to Reviewer oPU9 (1/4)

We thank Reviewer oPU9 for the detailed review and for recognizing RiboSBDD's motivation, cross-biomolecule transfer design, and benchmark curation. Below we address Q1–Q4 on conformational Δχ\Delta\chi validation, separate sampling ablations, OOD and lead-selection studies, and backbone transfer and atom-count priors.


Q1: Direct Conformational Evidence for Local Base Correction vs. Geometry Losses

Q1: Could the authors provide more direct evidence that Local Base Correction captures meaningful RNA conformational changes, such as angular prediction errors or comparisons with apo–holo structures? Could they also distinguish its contribution from that of the additional geometric losses?

Takeaway: Holding 81k updates, pψp_\psi, the frozen protein source, and the contact/steric loss schedule fixed across 713 pockets, Local Base Correction Δχ\Delta\chi delivers over half of the IMP% gain and nearly half of the Clash% reduction, and directly recovers glycosidic torsion shifts on synthetic perturbations and 14 split-disjoint SMARTFlexDB [11] Apo-to-Holo pairs.

  • Geometry losses vs. base rotation: In Table oPU9 Q1 01, rigid Lcontact+Lsteric\mathcal L_{\mathrm{contact}}+\mathcal L_{\mathrm{steric}} with Δχt≡0\Delta\chi_t \equiv 0 accounts for roughly half of the Clash% and IMP% improvement, while rotating nucleobases around χ\chi delivers the remaining half across training and Phase I inference at τ≥500\tau \ge 500.

Table oPU9 Q1 01: Separating geometry losses from local base flexibility across 713 pockets, matching Table 18.

MethodValid% ↑QED ↑SA ↑BA JS ↓Contact ↑Clash% ↓IMP% ↑Div. ↑
w/o Flex., diffusion only91.660.350.620.4900.745.8065.270.71
w/o Flex. + Geom. Losses, Δχ≡0\Delta\chi\equiv 091.830.340.620.4870.755.3867.420.70
w/o Infer. Rot., Δχ=0\Delta\chi=091.970.340.620.4850.755.1868.630.70
RiboSBDD Full92.220.340.620.4830.765.0070.120.70
  • Synthetic recovery & rotation schedule: Across N=713N=713 pockets at τ∈[500,999]\tau\in[500,999] in Table oPU9 Q1 02a, Δχ\Delta\chi cuts angular MAE by 71.2% at 91.6% sign accuracy. On the fixed benchmark pocket C\mathcal C in Table 19B, two-phase rotation at τ≥500\tau \ge 500 followed by freezing C\mathcal C at τ<500\tau < 500 reaches 5.00% Clash% and 70.12% IMP%, outperforming static Δχ≡0\Delta\chi \equiv 0 at 5.18%/68.63% and continuous τ≥0\tau \ge 0 at 5.14%/69.74%; scoring τ≥0\tau \ge 0 against the altered pocket C~0\widetilde{\mathcal C}_0 yields 4.82%/70.38% only because it modifies the evaluation receptor itself rather than scoring on C\mathcal C.
  • External Apo-to-Holo pairs: In Tables oPU9 Q1 02a–02b, we filter 186 SMARTFlexDB [11] pairs to 29 split-disjoint pairs and 14 ligand-free Apo clefts with Kabsch [1] backbone RMSD≤1.5\mathrm{RMSD} \le 1.5 Å across 58 active bases satisfying ∣Δχexp,r∣≥5.0∘|\Delta\chi_{\mathrm{exp},r}| \ge 5.0^\circ. Conditioned only on Capo\mathcal C_{\mathrm{apo}}, RiboSBDD predicts torsion shift signs at 74.1% accuracy, or 43/58 with pair-cluster CI [63.8%, 83.6%], pair sign-test p=0.0129p=0.0129, and binomial p=3.1×10−4p=3.1\times 10^{-4}, vs. 53.4% or 31/58 for canonical anti restoration toward χanti=−160∘\chi_{\mathrm{anti}}=-160^\circ; Holo Clash% and IMP% improve with 95% CIs [−6.68, −4.92%] and [+14.8, +21.8%].

Table oPU9 Q1 02a: Direct Δχ\Delta\chi recovery on synthetic perturbations across N=713N=713 pockets and SMARTFlexDB Apo–Holo across N=14N=14 pairs and 58 bases, matching Table 19A,C.

RegimeMetricUnrotated, Δχt≡0\Delta\chi_t\equiv 0Canonical antiRiboSBDD
Synth., τ≥500\tau \ge 500Angular MAE ↓6.38°---1.84°
Synth., τ≥500\tau \ge 500Angular RMSE ↓7.92°---2.41°
Synth., τ≥500\tau \ge 500Base RMSD ↓0.42 Å---0.13 Å
Synth., τ≥500\tau \ge 500Sign Accuracy ↑------91.6%
Apo–Holo, τ=500\tau=500Sign Accuracy ↑---53.4%: 31/5874.1%: 43/58
Apo–Holo, τ=500\tau=500Angular MAE ↓7.62°---5.18°
Apo–Holo, τ=500\tau=500Base RMSD ↓0.69 Å---0.49 Å

Table oPU9 Q1 02b: Apo-conditioned generation on Capo\mathcal C_{\mathrm{apo}} scored in the Holo frame across N=14N=14 pairs from Table 19C.

ConfigurationHolo Contact ↑Holo Clash% ↓Holo IMP% ↑
TargetDiff P→R, rigid Apo0.6812.04%43.85%
Rigid + Geom. Losses, rigid Apo0.717.52%55.12%
RiboSBDD Full, Apo + Δχ\Delta\chi0.746.24%62.14%

↳ Manuscript updates: Secs. 3.2–3.3, 4.3, 5, Table 1, App. B.4, D.2 (Table 19); separated geometry losses from Δχ\Delta\chi and added synthetic and SMARTFlexDB Apo–Holo tests.

AuthorsOpenReview ↗

Response to Reviewer oPU9 (2/4)

Q2: Coordinate Composition vs. Particle Resampling in Adaptive Sampling

Q2: How does the proposed adaptive sampling method compare with simpler alternatives, such as fixed-weight coordinate averaging? Could the contributions of coordinate composition and particle resampling be evaluated separately?

Takeaway: Holding our flexible RNA checkpoint, pψ(N∣C)p_\psi(N\mid\mathcal C), frozen protein source, and 125-step schedule fixed across 713 pockets, fixed source mixing, equal-density projection κt\kappa_t, and particle resampling each deliver distinct, additive gains in Table oPU9 Q2 01.

Table oPU9 Q2 01: Separate evaluation of fixed mixing, adaptive κt\kappa_t, and resampling across 713 pockets from Table 1 and Table 18.

SettingValid% ↑QED ↑SA ↑BA JS ↓Contact ↑Clash% ↓IMP% ↑Div. ↑
Target only: w/o Sampling89.590.300.590.4980.789.2060.310.67
Fixed mixing κ=0.5\kappa=0.5, no resampling91.420.330.610.4890.766.2465.380.71
Adaptive mixing κt\kappa_t, no resampling91.790.340.610.4860.765.5167.740.71
Fixed mixing κ=0.5\kappa=0.5 + resampling92.040.330.610.4860.765.4668.180.71
RiboSBDD Full: κt\kappa_t + resampling92.220.340.620.4830.765.0070.120.70
  • Stepwise mechanism: Static κ=0.5\kappa=0.5 mixing restores protein-pretrained steric avoidance by cutting Clash% from 9.20% to 6.24%; equal-density projection κt=clip(κt∗,−0.5,1.5)\kappa_t=\mathrm{clip}(\kappa_t^\ast,-0.5,1.5) from Skreta et al. [2] in Eq. 10 cancels shared noise along dt=μr−μsd_t=\mu_r-\mu_s on 94.3% of steps down to 5.51%; and Feynman-Kac resampling from Skreta et al. [3] in Eq. 11 with κˉt=clip(κt,0,1)\bar\kappa_t=\mathrm{clip}(\kappa_t,0,1) triggers when ESS<B/2\mathrm{ESS} < B/2 on 18.4% of Phase I steps, reaching 5.00% Clash% and 70.12% IMP% in 14.2 s vs. 9.46% and 56.42% for 315-NFE TD P→R at 14.3 s in Table oPU9 Q4 01.

↳ Manuscript updates: Secs. 3.3, 4.3, Table 1, App. B.6, D.2 (Table 18); evaluated fixed mixing, adaptive κt\kappa_t, and particle resampling separately.

AuthorsOpenReview ↗

Response to Reviewer oPU9 (3/4)

Q3: RNA Diversity, Neighboring Proteins, External Pairs, and Lead Selection

Q3: Could the authors provide further analyses of RNA target diversity, ligand interaction quality, and neighboring protein effects in RiboSBDDBench? Would evaluation on an independent dataset or downstream task be feasible?

Takeaway: RiboSBDD leads all evaluated baselines on Valid%, Contact, Clash%, and IMP% across four separately evaluated OOD subsets, all five RNA classes, full-assembly and 250-pocket protein-free checks, 14 external SMARTFlexDB [11] pairs, and post-hoc lead selection; prospective wet-lab binding assays remain future work.

  • OOD strata & neighboring proteins: In Table oPU9 Q3 01, across separately evaluated OOD subsets covering s<0.4s < 0.4 vs. the RNA+protein union, RNA-unseen scaffolds, RNA-unseen Rfam/Infernal [7, 8] families, and US-align TMpocket<0.50\mathrm{TM}_{\mathrm{pocket}} < 0.50, plus the 250 protein-free pockets and all five RNA classes in Table 24B where RiboSBDD reaches 61.2%–71.8% IMP% vs. 57.7%–65.7% for TD P and 47.5%–56.1% for TD P→R, RiboSBDD leads Valid% at 91.5%–93.1% and Contact at 0.75–0.77 while reaching 0.41–0.43 QED and 0.66–0.67 SA across the OOD and protein-free subsets. Furthermore, 521/713 pockets lack protein atoms within 5 Å in Table 9, and RiboSBDD achieves the lowest Clash% across all 15 configurations on pocket.pdb at 5.00%, full RNA at 5.07%, and RNA+Protein at 9.09% vs. 9.50% for TD P and 14.29% for TD P→R in Table 21.

Table oPU9 Q3 01: Clash% ↓ and IMP% ↑ across protein-free and OOD subsets, where DiffSBDD denotes P→R; full metrics in Tables 22–24.

SubsetDiffSBDD Clash↓DiffSBDD IMP↑TD P Clash↓TD P IMP↑TD P→R Clash↓TD P→R IMP↑RiboSBDD Clash↓RiboSBDD IMP↑
Protein-Free, N=250N=25058.14%46.58%5.63%61.18%9.64%52.18%4.86%68.43%
Tanimoto s<0.4s < 0.4, N=82N=8262.58%39.18%7.21%57.18%12.78%46.48%6.14%64.18%
RNA-Unseen Scaff., N=112N=11262.72%39.58%6.94%54.79%12.94%45.62%6.28%63.84%
RNA-Unseen Rfam, N=64N=6459.68%46.98%6.58%64.32%11.38%55.14%5.68%72.84%
TMpocket<0.50\mathrm{TM}_{\mathrm{pocket}} < 0.50, N=94N=9460.48%41.24%6.74%58.12%11.86%48.64%5.84%65.42%
  • External validation & lead ranking: Beyond the 14 SMARTFlexDB [11] pairs in Tables oPU9 Q1 02a–02b, Eq. 30 [4, 5] batch ranking under default <5% clash in Table oPU9 Q3 02a improves unselected Contact to 0.80 and independent LigandRNA [10] IMP% to 66.8% unranked and 82.6% Top-1 across all 589/713 parser-compatible pockets, including 79.7% on unseen Rfam.

Table oPU9 Q3 02a: Benchmark-wide selection under <5% clash across 713 pockets, 125 PDBs, and 589-pocket LigandRNA from Table 17A.

Method & TierValid%↑Pock. QED↑PDB QED↑SA↑Cont.↑Clash%↓Pock. IMP%↑PDB IMP%↑LigandRNA↑
DiffSBDD P→R Top-144.50%0.490.510.780.729.82%65.74%64.18%58.9%
TD P Top-188.58%0.530.550.750.741.65%80.15%78.64%73.8%
TD P→R Top-188.74%0.520.540.760.762.76%72.18%70.62%67.5%
RiboSBDD Top-392.22%0.520.540.730.791.43%85.16%83.42%---
RiboSBDD Top-192.22%0.560.580.770.801.18%88.64%87.12%82.6%
  • Riboswitch clash gates: On 6QN3 in Fig. 17c and Table oPU9 Q3 02b, we explicitly distinguish default <5%-clash Top-1 #13 at composite ranking score 0.610 from Top-1 under the strict 0%-clash filter #03 at score 0.559; on 3NPQ and 6FZ0, default <5% Top-1 already achieves 0.00% Clash%, so both gates select the identical pose.

Table oPU9 Q3 02b: Unseen-Rfam riboswitch Top-1 leads under default <5% versus strict 0% clash gates from Table 17B.

Target & RfamGateModel & PoseValid%↑QED↑SA↑Clash%↓Score ↓Crystal
6QN3, RF01704Default <5%TD P→R88.0%0.4420.700.00%−107.57−165.48
6QN3, RF01704Default <5%RiboSBDD #1394.0%0.3310.583.80%−241.32−165.48
6QN3, RF01704Strict 0%TD P→R88.0%0.4420.700.00%−107.57−165.48
6QN3, RF01704Strict 0%RiboSBDD #0394.0%0.4750.720.00%−180.15−165.48
3NPQ, RF01057<5% & 0%TD P→R88.0%0.4720.780.00%−142.34−218.62
3NPQ, RF01057<5% & 0%RiboSBDD92.0%0.5280.800.00%−234.86−218.62
6FZ0, RF01826<5% & 0%TD P→R90.0%0.4650.780.00%−162.58−289.44
6FZ0, RF01826<5% & 0%RiboSBDD94.0%0.5140.790.00%−329.88−289.44

↳ Manuscript updates: Secs. 4.2, 5, Tables 2a–2b, App. C.1–C.3 (Table 17), D.5 (Tables 21–24); added RNA diversity/OOD analyses, protein checks, and <5%/0% riboswitch ranking.

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

The authors proposed RiboSBDD, a RNA-conditioned ligand generation framework, together with an emerging benchmark dataset called RiboSBDDBench, which includes 8,006 entries assessed by various types of metrics. Compared with existing works, a unified atomic representation system across protein and RNA was applied. Specifically, the authors managed to conduct RNA specific fine tuning towards a pretrained model on protein-ligand complex data.

Critical strengths
  1. The authors introduced RiboSBDDBench, a promising benchmark dataset that might be used to conduct precision ligand design in future.
  2. A unified atomic representation system across protein and RNA was applied.
Critical weaknesses
  1. [Major] Significance of each composition: Though the ablation data is presented in Table 1 and Appendix D.2, it is quite confused for me to distinguish the specific contribution of each component in RiboSBDD. Specifically, would you please specify the significance of 'Sampling'? Additionally, as the authors mentioned, contact recovery (↑) might not yield better packing (lines 429-430). Then would you please document the necessity of including this metric? Another follow-up question is: Could the authors shortlist Top-5 most essential metrics that we should pay attention to, specifically in the downstream task of RNA conditioned ligand generation? Regarding QED, I assume it refers to "Quantitative Estimation of Drug-likeness" (0-1). Then how can RiboSBDD, as a SBDD pipeline, be applied if the generated ligands' QED are not that high, compared with the existing models (i.e., Table 1)?

  2. [Major] Dataset: According to the submission (i.e., lines 202-203), RiboSBDD would be firstly pretrained on protein-ligand complexes. Would the authors please disclose some more information of the (pretraining) protein-ligand complex dataset? For instance, how many source PDB entries have been retrieved? What are the detailed procedures of the filtering?

  3. [Major] Ranking of the generated ligand(s): I'm wondering how the candidatory ligands generated by RiboSBDD will be ranked. According to the figures of several qualitative cases, it seems that the first generated ligand's QED (↑) / energy score (↓) is not always the best among the pool. Does it mean that the users have to send lots of generation requests towards each pocket? Is it possible for RiboSBDD to generate the best candidate directly, after multiple rounds of inference?

Questions for authors

[Minor] Please consider proofreading the manuscript afterwards. Specifically, it is kind of misleading while reading such sentence, "Protein and RNA pockets shared a unified residue vocabulary of 20 amino acids and four canonical RNA nucleotides" (i.e., lines 153-154). I assume that the authors are trying to present the unify information across various types of heavy atoms. However, further corrections are necessary. Additionally, the full form of each abbreviation should be presented at the first presence.

11 comments
AuthorsOpenReview ↗

Response to Reviewer LoFT (1/5)

We thank Reviewer LoFT for the constructive review and for recognizing RiboSBDDBench and our unified atomic representation across proteins and RNA. Below we address Q1–Q8 on component contributions, adaptive sampling, metric complementarity, Top-5 metrics, downstream QED and lead ranking, pretraining provenance, single-batch runtime, and proofreading.


Q1: Individual Contribution of Each RiboSBDD Component

Q1: Significance of each composition, it is quite confused for me to distinguish the specific contribution of each component in RiboSBDD.

Takeaway: Each RiboSBDD module resolves a distinct physical bottleneck in narrow RNA grooves: boundary geometry losses supervise contact and repulsive distances, local base rotation Δχ\Delta\chi adapts pocket nucleobases, and adaptive sampling preserves 3D steric avoidance.

  • Isolating each module: Holding 81k updates, pψ(N∣C)p_\psi(N\mid\mathcal C), and 125 steps fixed across 713 pockets in Tables LoFT Q1 01–02, rigid Lcontact+Lsteric\mathcal L_{\mathrm{contact}}+\mathcal L_{\mathrm{steric}} within [2.2 Å, 6.0 Å] accounts for roughly half of the flexibility gain by cutting Clash% from 5.80% to 5.38%, base rotation Δχ\Delta\chi delivers the remaining half down to 5.00%, and adaptive sampling cuts Clash% from 9.20% to 5.00%.

Table LoFT Q1 01: Six controlled comparisons isolating each RiboSBDD component.

Controlled ComparisonFactors Held FixedIsolated Mechanism
Rigid + Geom. vs. w/o Flex.Rigid pocket, 81k, pψp_\psi, 125 steps, samplerContact & steric losses Lcontact+Lsteric\mathcal L_{\mathrm{contact}}+\mathcal L_{\mathrm{steric}}
RiboSBDD Full vs. Rigid + Geom.Contact/steric schedule, pψp_\psi, samplerBase rotation Δχ\Delta\chi and angular loss Lχ\mathcal L_\chi
RiboSBDD Full vs. infer Δχ=0\Delta\chi=0Flexibly trained checkpoint, samplerPhase I inference rotation Δχ\Delta\chi at τ≥500\tau \ge 500
Fixed Mixing vs. w/o SamplingFlexible checkpoint, pψp_\psi, 125 stepsFrozen protein source at κ=0.5\kappa=0.5
Adaptive vs. Fixed MixingSource & target checkpointsEqual-density projection κt\kappa_t, Eq. 10 [1]
Resampling On vs. OffMixing rule κ=0.5\kappa=0.5 or κt\kappa_tFeynman-Kac resampling, Eq. 11 [2]

Table LoFT Q1 02: Component-wise ablations across 713 pockets; all 11 metrics appear in Table 1 and Table 18.

SettingValid% ↑QED ↑SA ↑BA JS ↓Contact ↑Clash% ↓IMP% ↑Div. ↑
TargetDiff P→R Baseline88.740.340.680.5010.739.9054.320.74
w/o Flexibility, diffusion only91.660.350.620.4900.745.8065.270.71
Rigid + Geometry Losses91.830.340.620.4870.755.3867.420.70
RiboSBDD, infer Δχ=0\Delta\chi=091.970.340.620.4850.755.1868.630.70
w/o Sampling, κt≡1\kappa_t\equiv 189.590.300.590.4980.789.2060.310.67
Fixed Mixing κ=0.5\kappa=0.5, no resamp.91.420.330.610.4890.766.2465.380.71
Adaptive κt\kappa_t, no resamp.91.790.340.610.4860.765.5167.740.71
Fixed Mixing κ=0.5\kappa=0.5 + resamp.92.040.330.610.4860.765.4668.180.71
RiboSBDD Full: κt\kappa_t + resamp.92.220.340.620.4830.765.0070.120.70

↳ Manuscript updates: Secs. 3.2, 4.3, Table 1, App. D.2 (Table 18); added controlled component-isolation rows separating geometry losses, Δχ\Delta\chi, and sampling.


Q2: Significance and Mechanism of Adaptive "Sampling"

Q2: Would you please specify the significance of “Sampling”?

Takeaway: "Sampling" denotes our inference-time Adaptive Source and Target Sampling in Sec. 3.3, which combines 3D coordinate updates from the frozen protein source FθsF_{\theta_s} and RNA target FθrF_{\theta_r} to prevent steric clashes while preserving RNA chemical specificity.

  • Protein source prior: Across the five sampling rows sharing the identical RNA checkpoint in Table LoFT Q1 02, FθrF_{\theta_r} predicts ligand atom types while FθsF_{\theta_s} transfers all-atom 3D geometry via aja_j with the 4 RNA residue embeddings zeroed; static κ=0.5\kappa=0.5 mixing cuts Clash% from 9.20% to 6.24%.
  • Equal-density κt\kappa_t & resampling: Hyperplane projection [1] in Eq. 10 cancels shared noise along dt=μr−μsd_t=\mu_r-\mu_s on 94.3% of steps down to 5.51% Clash%, and Feynman-Kac resampling [2] in Eq. 11 with κˉt=clip(κt,0,1)\bar\kappa_t=\mathrm{clip}(\kappa_t,0,1) ensuring log⁡Gt≤0\log G_t \le 0 triggers when ESS<B/2\mathrm{ESS} < B/2 to prune clashing trajectories, reaching 5.00% Clash% and 70.12% IMP% in 14.2 s per pocket on one A100 vs. 41.8 s for TargetDiff and 45.3 s for DiffSBDD.

↳ Manuscript updates: Secs. 3.3, 4.2–4.3, Table 1, App. B.6, D.2 (Table 18); clarified Adaptive Source and Target Sampling and added stepwise sampling ablations.

AuthorsOpenReview ↗

Response to Reviewer LoFT (2/5)

Q3: Why Contact Recovery Is Necessary Alongside Clash%

Q3: contact recovery (↑) might not yield better packing (lines 429-430). Then would you please document the necessity of including this metric?

Takeaway: Contact and Clash% enforce complementary 3D constraints that neither verifies alone: Contact measures 4 Å recall of native recognition nucleotides via ∣Rp(m)∩Rpref∣/∣Rpref∣|\mathcal R_p(m)\cap\mathcal R_p^{\mathrm{ref}}| / |\mathcal R_p^{\mathrm{ref}}| in Eq. 24 following PepFlow [3], Pocket2Mol [10], CBGBench [15], and GerNA-Bind [7], whereas Clash% penalizes van der Waals overlap.

  • Without Contact: A model drifting into bulk solvent outside the groove could trivially score 0% Clash%.
  • Without Clash%: A model can inflate Contact by pushing atoms into receptor overlap, as w/o Sampling does at 0.78 Contact and 9.20% Clash%, whereas RiboSBDD jointly leads all 12 baselines at 0.76 Contact and 5.00% Clash%.

↳ Manuscript updates: Sec. 4.1, App. C.3 (Eq. 24); documented physical complementarity of Contact Recovery and Clash%.


Q4: Top-5 Essential Downstream Metrics for RNA Ligand Generation

Q4: Could the authors shortlist Top-5 most essential metrics that we should pay attention to, specifically in the downstream task of RNA conditioned ligand generation?

Takeaway: After gating on chemical validity via Valid%, we prioritize five downstream metrics in Table LoFT Q4 01: 1. AnnapuRNA IMP%, 2. Receptor Clash%, 3. Contact Recovery, 4. QED, and 5. SA.

  • Hierarchy & selection: Metrics 1–3 enforce 3D groove energetics and sterics, Metrics 4–5 measure 2D medicinal tractability, and Diversity tracks batch chemical spread. High unranked 2D scores cannot rescue colliding poses [5–7] such as DiffSBDD P→R at 57.97% Pool Clash%, whereas RiboSBDD's 92.22% Valid% and 5.00% Pool Clash% supply abundant valid poses so Top-1 selection reaches 0.56 QED, 0.77 SA, 1.18% Clash%, and 88.64% IMP%.

Table LoFT Q4 01: Top-5 metrics across 713 pockets comparing unranked Pool from Table 1 against Top-1 ranked leads from Table 17, where DiffSBDD denotes P→R and Pool Valid% is 44.50%, 88.74%, and 92.22%.

Priority & MetricRole in RNA SBDDDiffSBDD PoolTD P→R PoolRiboSBDD PoolDiffSBDD Top-1TD P→R Top-1RiboSBDD Top-1
1. AnnapuRNA IMP% ↑3D pose energetics [4]48.84%54.32%70.12%65.74%72.18%88.64%
2. Receptor Clash% ↓Groove steric fit57.97%9.90%5.00%9.82%2.76%1.18%
3. Contact Recovery ↑Native residue recall [3]0.660.730.760.720.760.80
4. Drug-likeness QED ↑Medicinal tractability [8]0.370.340.340.490.520.56
5. Synthetic Access. ↑Tractability, SAraw<6\mathrm{SA}_{\mathrm{raw}} < 60.750.680.620.780.760.77

↳ Manuscript updates: Secs. 4.1–4.2, Table 2a, App. C.3 (Table 16); added Top-5 RNA SBDD metric guide and unranked-vs.-ranked comparisons.

AuthorsOpenReview ↗

Response to Reviewer LoFT (3/5)

Q5: Unranked QED vs. Practical SBDD Lead Selection

Q5: Regarding QED, I assume it refers to "Quantitative Estimation of Drug-likeness" (0-1). Then how can RiboSBDD, as a SBDD pipeline, be applied if the generated ligands' QED are not that high, compared with the existing models (i.e., Table 1)?

Takeaway: QED measures Quantitative Estimate of Drug-likeness on [0,1][0,1] [8] and averages 0.34 in RiboSBDD's unranked pool, matching native crystal ligands at 0.34 on these exact 713 pockets because polar polyamine/aminoglycoside ribosomes form two thirds of the test set; reference-free ranking of one parallel B=50B=50 batch via Eq. 30 lifts RiboSBDD Top-1 QED to 0.56 Equal-Pocket and 0.58 Equal-PDB at 1.18% Clash%, surpassing all ranked diffusion baselines as well as unranked Pocket2Mol at 0.41 and GraphBP at 0.48.

  • Unranked QED & steric trade-off: In Table LoFT Q5 01a, Bickerton et al. [8] calibrated QED on oral drugs, whereas bioactive RNA ligands require polar H-bond donors for polyanionic grooves [16, 17]. Equal-PDB weighting lifts both crystal and RiboSBDD unranked QED to 0.39, with RiboSBDD reaching 0.42 on protein-free pockets and 0.41 on unseen Rfam; meanwhile, Pocket2Mol and GraphBP reach 0.41 and 0.48 only at 16.68% and 80.01% Clash%, yielding 46.84% and 9.88% <5%-clash poses per request vs. 72.40% for RiboSBDD.

Table LoFT Q5 01a: Unranked ensembles across 713 pockets showing Equal-Pocket macro-means from Tables 13–15, versus pooled Table 1 in Q4, alongside Equal-PDB QED and <5%-clash request yield.

Unranked ConfigurationPool Valid% ↑Pocket QED ↑PDB QED ↑Pocket SA ↑Clash% ↓<5%-Clash / Req. ↑Contact ↑IMP% ↑
Native Crystal Ligands100.00%0.340.390.71------1.00---
Pocket2Mol P→R88.52%0.410.440.6416.68%46.84%0.7229.02%
GraphBP P→R80.89%0.480.490.6780.01%9.88%0.2526.58%
DiffSBDD P→R44.50%0.370.410.7557.82%14.20%0.6648.62%
TargetDiff P88.58%0.350.390.645.76%64.13%0.7063.81%
TargetDiff P→R88.74%0.340.390.689.94%61.76%0.7354.38%
RiboSBDD Full92.22%0.340.390.625.02%72.40%0.7670.18%
  • Post-hoc batch selection: In Table LoFT Q5 01b, ranking one B=50B=50 batch without reference ligands via Eq. 30 [5, 11–14] yields at least 3 passing poses in all 713 pockets without fallback. Beyond in-objective AnnapuRNA IMP%, RiboSBDD Top-1 leads outperform all ranked baselines on unselected Contact at 0.80, Equal-PDB Clash% and IMP% at 1.24% and 87.12%, and independent LigandRNA [18] win rate at 82.6% across all 589/713 parser-compatible pockets.

Table LoFT Q5 01b: Post-hoc ranked diffusion leads from one B=50B=50 batch under Eq. 30, matching Appendix C.3 Table 17.

Ranked Diffusion MethodTierPocket QED ↑PDB QED ↑Pocket SA ↑Clash% ↓Unsel. Contact ↑AnnapuRNA IMP% ↑LigandRNA IMP% ↑
DiffSBDD P→RTop-10.490.510.789.82%0.7265.74%58.9%
TargetDiff PTop-10.530.550.751.65%0.7480.15%73.8%
TargetDiff P→RTop-10.520.540.762.76%0.7672.18%67.5%
RiboSBDD FullTop-30.520.540.731.43%0.7985.16%---
RiboSBDD FullTop-10.560.580.771.18%0.8088.64%82.6%

↳ Manuscript updates: Fig. 1, Secs. 4.1–4.2, Table 2a, App. C.3 (Table 17); added native crystal QED baselines, Eq. 30 ranking protocol, and LigandRNA validation.

Reviewer_79NTOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Can you compare BindGRPO with at least one simpler post-training baseline using the same initial model, training targets, and reward-evaluation budget? Which policy-surrogate choices materially affect downstream success?

Does the two-step AlphaFold3 reward preserve rankings relative to full-step evaluation for co-designed sequences sampled during and after RL, including the unregularized run? This would help distinguish reward-approximation exploitation from other causes of the reward–success mismatch.

How were the final checkpoint and KL coefficient selected? Were Cao benchmark results consulted during these choices, and are the reported gains reproducible across independent training runs?

How does post-training affect sequence diversity, backbone diversity, and the number of distinct successful designs? Please also report the computational cost of target screening and RL training so that the practical trade-off can be assessed.

Critical strengths

A useful benchmark and systematic transfer-learning study. RNA-conditioned small-molecule generation is a worthwhile setting with limited structural data. The dataset curation, grouped splitting, and comparison of protein-only, RNA-only, and transfer-learning regimes across four architectures provide a useful experimental foundation. Grouping shared PDB entries and RNA90 clusters is preferable to randomly splitting individual complex instances. The consistent improvements in bond geometry and contact recovery from protein pretraining strengthen the transfer-learning motivation.

Complementary evaluation reveals meaningful strengths and trade-offs. The paper evaluates chemical descriptors, local geometry, receptor clashes, contact recovery, and RNA-specific pose scores rather than relying on a single metric. Relative to fine-tuned TargetDiff, RiboSBDD reduces the atom-level clash rate from 9.90% to 5.00% and increases the AnnapuRNA improvement rate from 54.32% to 70.12%. The authors also report that synthetic accessibility and within-pocket diversity decrease, and correctly avoid treating improved pose scores as evidence of experimentally confirmed binding.

Critical weaknesses

The evaluation does not yet establish comparable end-to-end performance across methods. Appendix C.3 explains that failed reconstruction and evaluation are excluded, and that Contact and IMP pool only successfully evaluated molecules. Consequently, methods can be compared on different subsets of outputs and effectively different pocket weightings despite receiving the same number of generation requests. The manuscript defines accepted-fragment validity and evaluation coverage, but does not report the corresponding results alongside the main comparisons. Without these denominators, higher conditional scores need not imply more useful candidates per request. There is also an important receptor-context issue: the “Clean” criterion permits up to two-thirds of nearby RNA-plus-protein heavy atoms to be protein atoms, while the model input and clash evaluation omit those proteins. Generated molecules could therefore occupy space that is unavailable in the deposited complex. The prevalence and effect of this issue should be quantified.

The ablations do not isolate the claimed contributions of flexibility and adaptive sampling. The “w/o Flexibility” control removes not only base perturbation and correction, but also the contact and steric losses. Since these losses directly target the reported interface metrics, the improvement cannot be attributed specifically to learning nucleobase adjustments. This distinction is particularly relevant because inference starts from the observed pocket at every step and returns to that pocket for low-noise denoising. Similarly, “w/o Sampling” simultaneously removes source guidance, adaptive composition, and resampling, leaving their individual contributions unresolved. The comparison with fine-tuned TargetDiff additionally changes the ligand-size policy from an RNA training histogram to a learned pocket-conditioned estimator and changes the reverse schedule. These differences could affect contact recovery and clashes independently of the proposed modules. The existing controls support the combined system, but provide weaker evidence for its individual mechanisms.

Questions for authors

Please report reconstruction, sanitization, accepted-fragment, and scoring coverage for every method, together with useful-output rates per generation request. Do the main rankings persist under equal-pocket aggregation and uncertainty estimates that respect the grouped structure of the test set?

How many test pockets contain nearby protein atoms, and how do the results change on a strictly protein-free subset or when checking clashes against the complete deposited receptor environment?

Can you provide a rigid-pocket control retaining the same contact and steric losses, continuation budget, atom-count predictor, and reverse schedule? Separately, how much do source guidance, adaptive mixing, and particle resampling each contribute relative to simpler fixed-weight mixing?

What RNA families, ligand identities, and ligand scaffolds are shared across the training and test partitions? RNA90 grouping prevents near-identical sequence overlap but does not establish family-level or pocket-level separation. Reporting performance by test-to-training similarity would clarify the benchmark’s generalization scope.

6 comments
AuthorsOpenReview ↗

Response to Reviewer 79NT (1/4)

We thank Reviewer 79NT for the rigorous review and for recognizing RiboSBDDBench and our cross-biomolecule transfer framework. Below we address Q1–Q4 on pipeline funnel and grouped uncertainty, neighboring protein checks, controlled module ablations, and stratified generalization.

Q1: Pipeline Funnel, Per-Request Yields, and Group-Aware Uncertainty

Q1: Please report reconstruction, sanitization, accepted-fragment, and scoring coverage for every method, together with useful-output rates per generation request. Do the main rankings persist under equal-pocket aggregation and uncertainty estimates that respect the grouped structure of the test set?

Takeaway: We score 100% of accepted molecules across all 35,650 requests with zero pose drops, and RiboSBDD leads all 12 baselines on Valid%, Contact, Clash%, and IMP% across per-request yields and Equal-Pocket, Equal-PDB, and Equal-Group means.

  • 100% scoring coverage: Under SBDD rules [1–4], OpenBabel bond recovery, RDKit sanitization, and the ≥⌊0.8N⌋\ge \lfloor 0.8N \rfloor fragment filter define G\mathcal G in Table 79NT Q1 01. All 713 pockets have native contacts and finite AnnapuRNA [5] scores, so ∣G∣=∣Gcontact∣=∣Gscore∣|\mathcal G| = |\mathcal G^{\mathrm{contact}}| = |\mathcal G^{\mathrm{score}}| for all 15 models; Novelty uses np≥1n_p \ge 1 and Diversity np≥2n_p \ge 2.
  • Per-request yields: On pocket.pdb via Eq. 22, RiboSBDD reaches 72.40% <5%-clash yield, or 78.51% of G\mathcal G, and 44.70% 0%-clash yield in Fig. 3b, or 48.47% of G\mathcal G, vs. 43.91/47.91% w/o Flex. and 28.11/31.38% w/o Samp. DiffSBDD P reaches 63.90% conditional IMP% only on the 21.42% surviving reconstruction, sanitization, and the fragment filter, giving 13.69% IMP/Req. vs. 64.66% RiboSBDD; <5%-clash IMP/Req. is 52.42% vs. 40.91% TD P, 34.09% TD P→R, 5.62% DiffSBDD P→R, and 1.80% DiffSBDD P.

Table 79NT Q1 01: Funnel counts and per-request yields across 35,650 requests; all % divide by 35,650, Table 13.

RegimeModelRecon.Sanit.Accepted=ScoredValid/Req.↑<5% Count↑<5%/Req.↑IMP/Req.↑
PTargetDiff P35,65035,58431,57988.58%22,86364.13%56.57%
PDiffSBDD P35,64834,6797,63721.42%2,2996.45%13.69%
P→RTargetDiff P→R35,65035,56131,63688.74%22,01961.76%48.20%
P→RDiffSBDD P→R35,64834,90615,86444.50%5,06114.20%21.73%
P→RPocket2Mol P→R35,65035,09431,55788.52%16,69846.84%25.74%
P→RGraphBP P→R35,64835,40728,83880.89%3,5229.88%21.45%
OursRiboSBDD35,65035,59132,87692.22%25,81272.40%64.66%
Ablat.w/o Flex.35,65035,57432,67791.66%24,90269.85%59.83%
Ablat.w/o Samp.35,65035,51231,93989.59%22,50963.14%54.03%
  • Grouped means & CIs: In Table 79NT Q1 02, RiboSBDD yields 46.11±3.2446.11 \pm 3.24 accepted poses and 43.8±3.643.8 \pm 3.6 unique SMILES per pocket, within 0.06 points of pooled Table 1 means. Equal-PDB and Equal-Group weight polyamine/aminoglycoside ribosomes [10] equally, lifting QED and SA while RiboSBDD leads on Valid%, Contact, Clash%, and IMP%. In 125-PDB and 14-group bootstraps, RiboSBDD outperforms TD P→R in 14/14 groups at p=1.2×10−4p=1.2\times 10^{-4} with 14-group Clash% and IMP% CIs [−5.68, −4.56%] and [+14.1, +17.3%], surpasses TD P in 13/14 and 14/14 groups with [−1.18, −0.34%] and [+4.68, +7.92%], and surpasses DiffSBDD P in 14/14 and 13/14 groups with [+4.12, +8.38%] IMP% CI and +50.97% IMP/Req.

Table 79NT Q1 02: Pocket 713 ±SD, PDB 125 ±SEM, and Group 14 ±SEM means; 15 models in Tables 14–15.

TierModelValid% ↑QED ↑SA ↑Contact ↑Clash% ↓IMP% ↑Div. ↑
PocketTD P88.58±7.620.35±0.100.64±0.100.70±0.145.76±6.7163.81±16.880.85±0.10
PocketTD P→R88.74±7.810.34±0.100.68±0.100.73±0.149.94±8.4354.38±17.660.74±0.10
PocketDiffSBDD P→R44.50±13.110.37±0.100.75±0.100.66±0.1457.82±14.0748.62±18.030.83±0.10
PocketRiboSBDD92.22±6.480.34±0.100.62±0.100.76±0.105.02±6.2470.18±15.560.70±0.10
PDBTD P88.72±0.540.39±0.010.66±0.010.70±0.016.12±0.4262.14±1.520.85±0.01
PDBTD P→R88.91±0.550.39±0.010.70±0.010.73±0.0110.18±0.5853.84±1.640.74±0.01
PDBRiboSBDD92.48±0.440.39±0.010.66±0.010.76±0.014.98±0.3669.42±1.380.71±0.01
GroupTD P88.64±0.920.39±0.020.66±0.020.70±0.026.18±0.7461.92±2.480.85±0.01
GroupTD P→R88.82±0.950.39±0.020.70±0.020.73±0.0210.22±0.9653.68±2.620.74±0.01
GroupRiboSBDD92.51±0.780.39±0.020.66±0.020.76±0.024.96±0.6269.38±2.240.71±0.01

↳ Manuscript updates: Secs. 4.1–4.2, Tables 1 & 2a, App. C.3 (Tables 13–15); added Valid%, funnel counts, <5%-clash yields, and grouped means.

AuthorsOpenReview ↗

Response to Reviewer 79NT (2/4)

Q2: Neighboring Protein Context, Full-Assembly Clashes, and Protein-Free Subset

Q2: How many test pockets contain nearby protein atoms, and how do the results change on a strictly protein-free subset or when checking clashes against the complete deposited receptor environment?

Takeaway: Across the 713 test pockets, 73.07% have zero protein heavy atoms within 5 Å and 35.06% come from strictly protein-free PDB entries; RiboSBDD achieves the lowest Clash% across all 12 baselines on complete deposited RNA at 5.07% and RNA+Protein at 9.09%, and leads all 12 baselines on Valid%, BA JS, Contact, Clash%, and IMP% across the 250 protein-free pockets.

  • Protein proximity: In Table 79NT Q2 01, over two thirds of test pockets contain no nearby protein atoms, and 250 pockets come from PDB entries with zero protein chains anywhere in the deposited assembly.

Table 79NT Q2 01: Protein presence across the 713 Clean test pockets, matching Appendix C.1 Table 9.

CriterionPocket CountShareStructural Meaning
Protein heavy atoms within 5 Å of native ligand192 / 71326.93%Local ribonucleoprotein proximity
Zero protein heavy atoms within 5 Å of native ligand521 / 71373.07%No nearby protein atoms
Zero protein chains anywhere in deposited PDB250 / 71335.06%Strictly protein-free PDB complex
  • Complete deposited assemblies: In Table 79NT Q2 02, moving from pocket.pdb, which retains explicit receptor hydrogens when present, to heavy-atom Full Deposited RNA changes RiboSBDD by only +0.07 pp from 5.00% to 5.07%; the other pocket-contained diffusion and Pocket2Mol models shift downward by 0.24–0.58 pp, whereas GraphBP rises by 1.66–5.47 pp as poses protrude into surrounding helices. Adding unmodeled ribosomal proteins in Full RNA+Protein leaves RiboSBDD with the lowest Clash% at 9.09% across all 15 configurations.

Table 79NT Q2 02: Atom-level Clash% ↓ across three receptor scopes on all 713 pockets; all 15 rows appear in Table 21.

RegimeModelValid% ↑10 Å RNA Pocket ↓Full Deposited RNA ↓Full RNA + Protein ↓
PTargetDiff P88.58%5.73%5.27%9.50%
P→RTargetDiff P→R88.74%9.90%9.61%14.29%
P→RDiffSBDD P→R44.50%57.97%57.56%58.76%
P→RPocket2Mol P→R88.52%16.74%16.23%17.14%
P→RGraphBP P→R80.89%79.97%85.44%87.05%
OursRiboSBDD Full92.22%5.00%5.07%9.09%
Ablat.w/o Flexibility91.66%5.80%5.45%9.38%
Ablat.w/o Sampling89.59%9.20%8.96%13.42%
  • Strictly protein-free benchmark: Across the 250 protein-free pockets in Table 79NT Q2 03, RiboSBDD rises to 0.42 QED and 0.67 SA and leads all 12 baselines on Valid%, BA JS, Contact, Clash%, and IMP%, while w/o Sampling trades 9.16% Clash% for 0.78 Contact.

Table 79NT Q2 03: Key models on the 250 strictly protein-free pockets, pooled as in Table 1; all 15 rows appear in Table 22.

SettingValid% ↑QED ↑SA ↑BA JS ↓Contact ↑Clash% ↓IMP% ↑Div. ↑
TargetDiff P89.180.410.680.500.715.6361.180.85
TargetDiff P→R89.380.420.710.490.749.6452.180.75
DiffSBDD P→R45.180.440.770.590.6658.1446.580.84
Pocket2Mol P→R89.080.450.670.530.7316.4928.610.77
GraphBP P→R81.460.490.690.590.2580.1425.430.86
RiboSBDD Full93.140.420.670.480.774.8668.430.71
w/o Flexibility92.580.430.670.490.755.6863.180.72
w/o Sampling90.440.390.640.500.789.1658.140.68

↳ Manuscript updates: Secs. 4.1–4.2, Table 2b, App. C.1 (Table 9), D.5 (Tables 21–22); added 5 Å protein proximity audit, full-assembly Clash%, and protein-free subset evaluation.

AuthorsOpenReview ↗

Response to Reviewer 79NT (3/4)

Q3: Rigid-Pocket Geometry Loss Control and Adaptive Sampling Decomposition

Q3: Can you provide a rigid-pocket control retaining the same contact and steric losses, continuation budget, atom-count predictor, and reverse schedule? Separately, how much do source guidance, adaptive mixing, and particle resampling each contribute relative to simpler fixed-weight mixing?

Takeaway: Holding the 81k update budget, pψ(N∣C)p_\psi(N\mid\mathcal C), frozen protein source, and 125-step schedule fixed, rigid geometry losses, base rotation Δχ\Delta\chi, protein-source mixing, equal-density projection κt\kappa_t, and particle resampling each deliver statistically significant improvements under both 125-PDB and 14-group paired cluster-bootstrap CIs.

  • Schedule & compute alignment: In Table 79NT Q3 01, matching TargetDiff P→R to 81k updates, 125 steps, pψp_\psi, or 315 NFE at 14.3 s vs. RiboSBDD's 14.2 s per pocket on one A100 shifts Clash% by at most 0.44 pp and IMP% by at most 2.10 points.
  • Flexibility vs. geometry losses: Across Tables 79NT Q3 01–02, rigid Lcontact+Lsteric\mathcal L_{\mathrm{contact}}+\mathcal L_{\mathrm{steric}} with Δχ≡0\Delta\chi \equiv 0 accounts for 52.5% of the Clash% reduction and 44.3% of the IMP% gain, while glycosidic rotation Δχ\Delta\chi [13, 14] accounts for the remaining 47.5% and 55.7%; Phase I inference rotation alone contributes −0.18% Clash% across 12/14 groups and +1.49% IMP% across 13/14 groups.
  • Isolating adaptive sampling: Fixed κ=0.5\kappa=0.5 mixing cuts Clash% by −2.96% and lifts IMP% by +5.07% across 14/14 groups; equal-density projection κt\kappa_t [6] cancels shared noise along dt=μr−μsd_t=\mu_r-\mu_s on 94.3% of steps; Feynman-Kac resampling [7] triggers when ESS<B/2\mathrm{ESS} < B/2 on 18.4% of Phase I steps to prune clashing trajectories.

Table 79NT Q3 01: Controlled flexibility, alignment, and sampling ablations across 713 pockets, pooled as in Table 1 and Appendix D.2 Table 18.

SettingValid% ↑QED ↑SA ↑BA JS ↓Contact ↑Clash% ↓IMP% ↑Div. ↑
TargetDiff P→R, Original: 75k / 1000, Hist. NN88.740.340.680.5010.739.9054.320.74
Matched TargetDiff P→R: 81k / 125, pψp_\psi89.120.340.660.4980.749.5256.180.71
Compute-Matched TD P→R: 81k / 315 NFE, pψp_\psi89.200.340.660.4970.749.4656.420.71
w/o Flexibility, diffusion loss only91.660.350.620.4900.745.8065.270.71
Rigid + geometry loss control, Δχ≡0\Delta\chi\equiv 091.830.340.620.4870.755.3867.420.70
RiboSBDD, flexible train, infer Δχ=0\Delta\chi=091.970.340.620.4850.755.1868.630.70
w/o Sampling, Target only89.590.300.590.4980.789.2060.310.67
Fixed mixing κ=0.5\kappa = 0.5, no resampling91.420.330.610.4890.766.2465.380.71
Adaptive mixing κt\kappa_t, no resampling91.790.340.610.4860.765.5167.740.71
Fixed mixing κ=0.5\kappa = 0.5 + particle resampling92.040.330.610.4860.765.4668.180.71
RiboSBDD, histogram count: 81k / 12591.580.350.670.4860.755.3667.180.75
RiboSBDD Full92.220.340.620.4830.765.0070.120.70

Table 79NT Q3 02: Incremental module gains under fixed 81k updates, 125 steps, and pψp_\psi, with 125-PDB and 14-group paired cluster-bootstrap 95% CIs from Table 18B.

Isolated Module StepΔ\DeltaClash% ↓125-PDB / 14-Group CIΔ\DeltaIMP% ↑125-PDB / 14-Group CIGroup Wins: Clash / IMP
+ Geom. Losses vs. w/o Flex.−0.42%[−0.66, −0.19] / [−0.72, −0.14]+2.15%[+1.14, +3.18] / [+0.96, +3.38]13/14, 13/14
+ Base Rotation Δχ\Delta\chi vs. Rigid + Geom.−0.38%[−0.61, −0.16] / [−0.67, −0.11]+2.70%[+1.64, +3.76] / [+1.42, +3.98]13/14, 14/14
+ Projection κt\kappa_t vs. Fixed κ=0.5\kappa=0.5−0.73%[−1.01, −0.46] / [−1.09, −0.39]+2.36%[+1.34, +3.38] / [+1.16, +3.58]13/14, 13/14
+ Resampling vs. Adaptive κt\kappa_t only−0.51%[−0.76, −0.27] / [−0.83, −0.21]+2.38%[+1.41, +3.36] / [+1.22, +3.54]13/14, 14/14

↳ Manuscript updates: Secs. 3.2–3.3, 4.3, Table 1, App. B.6, D.2 (Table 18); added rigid geometry-loss control, sampling decomposition, and cluster-bootstrap CIs.

Comments on the paper

4 comments
AuthorsOpenReview ↗

Overall Rebuttal (1/3)

We thank the Area Chair and Reviewers 79NT, LoFT, and oPU9 for their constructive feedback. Below we summarize our revisions and new experimental evidence across six themes; full tables appear in each reviewer response, and manuscript additions appear in blue across the 10-page main text and Appendices A–D.


1. Controlled Module Ablations, Backbone Transfer & Size Prior — Reviewer 79NT Q3; Reviewer LoFT Q1–Q2; Reviewer oPU9 Q1–Q2, Q4

Takeaway: Under matched 81k updates, 125 steps or 315 NFE, and atom-count prior pψ(N∣C)p_\psi(N\mid\mathcal C), RiboSBDD cuts Clash% from 9.52% to 5.00% and lifts AnnapuRNA [7] IMP% from 56.18% to 70.12%, with every module significant under 125-PDB and 14-group cluster bootstraps.

  • Schedule & compute alignment: Matching TargetDiff P→R [3] to 81k updates, 125 steps, pψp_\psi, or 315 NFE at 14.3 s vs. RiboSBDD's 14.2 s per pocket on one A100 shifts Clash% by ≤0.44\le 0.44 pp and IMP% by ≤2.10\le 2.10 points across Rows 1–2c.
  • Flexibility vs. geometry losses: Rigid contact and steric losses deliver half of the 3D gain at −0.42% Clash% and +2.15% IMP% in Row 5, while base rotation Δχ\Delta\chi [2] delivers the other half at −0.38% Clash% and +2.70% IMP% in Row 11.
  • Adaptive sampling & backbone transfer: Source mixing, equal-density projection κt\kappa_t [5], and Feynman-Kac resampling [6] each add distinct gains across Rows 3, 7–8, and 11, and transfer directly to DiffSBDD [4] by cutting Clash% from 57.97% to 38.62% and lifting IMP% from 48.84% to 60.75% in Row 14.
  • Atom-count prior & size-matched scoring: Under a shared histogram prior in Row 10, RiboSBDD retains 0.67 SA and 0.75 Diversity at 5.36% Clash% and 67.18% IMP%; RiboSBDD Full maintains a +14.8 point IMP% lead at Nheavy∈[18,20]N_{\mathrm{heavy}}\in[18,20] and improves proxy LE=sp/Nheavy\mathrm{LE}=s_p/N_{\mathrm{heavy}} to −9.96 vs. −8.83 for TD P→R and −8.62 for crystal ligands in Table oPU9 Q4 02.

Table G1: Bridge ablations and backbone transfer across 713 pockets; full 16-row bridge in App. D.2 Table 18 and Table 1.

ConfigurationSched. & PriorValid% ↑QED ↑SA ↑BA JS ↓Contact ↑Clash% ↓IMP% ↑Div. ↑
1. Original TargetDiff P→R75k/1000, Hist.88.740.340.680.5010.739.9054.320.74
2b. Matched TargetDiff P→R81k/125, pψp_\psi89.120.340.660.4980.749.5256.180.71
2c. Compute-Matched, 315 NFE81k/315, pψp_\psi89.200.340.660.4970.749.4656.420.71
3. w/o Sampling81k/125, pψp_\psi89.590.300.590.4980.789.2060.310.67
4. w/o Flexibility81k/125, pψp_\psi91.660.350.620.4900.745.8065.270.71
5. Rigid + Geometry Losses81k/125, pψp_\psi91.830.340.620.4870.755.3867.420.70
6. RiboSBDD, infer Δχ=0\Delta\chi=081k/125, pψp_\psi91.970.340.620.4850.755.1868.630.70
7. Fixed Mixing, no resamp.81k/125, pψp_\psi91.420.330.610.4890.766.2465.380.71
8. Adaptive Mixing, no resamp.81k/125, pψp_\psi91.790.340.610.4860.765.5167.740.71
10. RiboSBDD, Hist. count81k/125, Hist.91.580.350.670.4860.755.3667.180.75
11. RiboSBDD Full81k/125, pψp_\psi92.220.340.620.4830.765.0070.120.70
12. DiffSBDD P→R Baseline75k/1000, Hist.44.500.370.750.5900.6657.9748.840.83
14. RiboSBDD on DiffSBDD81k/125, pψp_\psi51.280.360.680.5640.7138.6260.750.78

↳ Manuscript updates: Secs. 3.2–4.3, Table 1, App. B.1–B.6, D.2 (Tables 18, 20); added controlled module ablations, DiffSBDD transfer, and atom-count diagnostics.


2. Direct Structural & Paired Apo–Holo Validation of Δχ\Delta\chi — Reviewer oPU9 Q1; Reviewer 79NT Q3

Takeaway: Local Base Correction Δχ\Delta\chi recovers nucleobase torsion shifts on synthetic perturbations and 14 split-disjoint SMARTFlexDB [12] Apo-to-Holo pairs, improving Holo Contact, Clash%, and IMP%.

  • Synthetic & schedule diagnostics: Across N=713N=713 pockets at τ∈[500,999]\tau\in[500,999], Δχ\Delta\chi cuts angular MAE and base RMSD by 71% and 69% at 91.6% sign accuracy; rotating at τ≥500\tau \ge 500 and freezing C\mathcal C at τ<500\tau < 500 outperforms static and continuous schedules.
  • External Apo-to-Holo pairs: Filtering 186 SMARTFlexDB [12] pairs to 29 disjoint pairs and 14 ligand-free Apo clefts with Kabsch backbone RMSD≤1.5\mathrm{RMSD} \le 1.5 Å yields 58 bases with ∣Δχexp,r∣≥5.0∘\lvert\Delta\chi_{\mathrm{exp},r}\rvert \ge 5.0^\circ. Apo-conditioned RiboSBDD reaches 74.1% sign accuracy, or 43/58 with cluster CI [63.8%, 83.6%] and p=0.0129p=0.0129, vs. 53.4% or 31/58 for canonical anti restoration, improving Holo Contact to 0.74, Clash% to 6.24%, and IMP% to 62.14%.

↳ Manuscript updates: Secs. 3.2–3.3, 4.3, 5, App. B.4, D.2 (Table 19); added synthetic Δχ\Delta\chi errors, schedule ablations, and SMARTFlexDB tests.

AuthorsOpenReview ↗

Overall Rebuttal (2/3)

3. Pipeline Funnel, Per-Request Yields & Group-Aware Inference — Reviewer 79NT Q1; Reviewer LoFT Q3–Q5, Q7; Reviewer oPU9 Q3

Takeaway: We score 100% of accepted molecules with zero pose drops, and RiboSBDD leads all 12 baselines on Valid%, Contact, Clash%, and IMP% across per-request yields and Equal-Pocket, Equal-PDB, and Equal-Split-Group means.

  • Per-request yields: RiboSBDD delivers 72.40% <5%-clash valid poses and 64.66% per-request IMP across all 35,650 requests in Table 79NT Q1 01, whereas DiffSBDD P reaches high conditional IMP% only on the 21.42% of requests surviving reconstruction, sanitization, and the fragment filter for 13.69% per-request IMP.
  • Cluster-aware statistics: Under 125-PDB and 14-group paired cluster bootstraps in Table 79NT Q1 02, RiboSBDD outperforms TargetDiff P→R in 14/14 split groups at p=1.2×10−4p=1.2\times 10^{-4} and surpasses TargetDiff P and DiffSBDD P in 13–14/14 groups, while Equal-PDB weighting raises RiboSBDD mean QED from 0.34 to 0.39 and SA from 0.62 to 0.66 once polar polyamine/aminoglycoside ribosomes [1] receive unit weight.

↳ Manuscript updates: Secs. 4.1–4.2, Tables 1 & 2a, App. C.3 (Tables 13–16); added stagewise funnel counts, <5%-clash yields, and grouped means with cluster CIs.


4. Neighboring Protein Context & Strictly Protein-Free Benchmark — Reviewer 79NT Q2; Reviewer oPU9 Q3

Takeaway: Across 713 test pockets, 521 or 73.07% have zero protein heavy atoms within 5 Å of the ligand; RiboSBDD achieves the lowest Clash% across all 12 baselines on full deposited RNA at 5.07% and RNA+Protein at 9.09%, and leads all 12 baselines on Valid%, BA JS, Contact, Clash%, and IMP% across the 250 strictly protein-free pockets.

↳ Manuscript updates: Secs. 4.1–4.2, Table 2b, App. C.1 (Table 9), D.5 (Tables 21–22); added 5 Å protein proximity audit, full-assembly Clash%, and 250-pocket protein-free results.


5. Stratified OOD Generalization & Batch Lead Ranking — Reviewer 79NT Q4; Reviewer LoFT Q4–Q5, Q7; Reviewer oPU9 Q3

Takeaway: RiboSBDD leads all evaluated baselines on Valid%, Contact, Clash%, and IMP% across four OOD strata and all five RNA classes, while one 14.2 s A100 batch selects high-QED, low-clash leads validated by LigandRNA [10].

  • OOD generalization: Across separately evaluated OOD subsets covering s<0.4s < 0.4 Tanimoto vs. RNA+protein ligands at N=82N=82, RNA-unseen scaffolds at N=112N=112, RNA-unseen Rfam/Infernal [8, 9] families at N=64N=64, and US-align [11] max⁡TMpocket<0.50\max\mathrm{TM}_{\mathrm{pocket}} < 0.50 at N=94N=94, plus five RNA classes, RiboSBDD holds +15.3 to +18.2 point IMP% margins over TD P→R and +5.8 to +9.1 over TD P except ribozymes at +13.7 and +3.3 in Tables 79NT Q4 02a–02b and Table oPU9 Q3 01.
  • Single-batch ranking: Ranking one B=50B=50 batch via Eq. 30 lifts RiboSBDD Top-1 QED to 0.56 Equal-Pocket and 0.58 Equal-PDB at 1.18% Clash%, 0.80 Contact, and 82.6% Top-1 win rate under LigandRNA [10] across all 589/713 parser-compatible pockets in Tables LoFT Q5 01a–01b and Tables oPU9 Q3 02a–02b; wet-lab binding assays remain future work.

↳ Manuscript updates: Secs. 4.2, 5, Tables 2a–2b, App. C.2 (Table 12), C.3 (Table 17), D.5 (Tables 23–24); added split overlap audit, OOD strata, and Eq. 30 lead ranking.

AuthorsOpenReview ↗

Overall Rebuttal (3/3)

6. Pretraining Provenance, Notation & Proofreading — Reviewer LoFT Q6, Q8

Takeaway: All P and P→R baselines and RiboSBDD share the exact same CrossDocked2020 v1.1 split comprising 100,000 training pairs from 15,153 PDBs and 100 disjoint validation pockets, and we expand every acronym at first use while separating atomic features aja_j from residue identities sjs_j within the 10-page main text.

↳ Manuscript updates: Abstract, Secs. 1–5, App. B.1–B.2 (Tables 4–5), C.3 (Eq. 29); documented CrossDocked2020 provenance, expanded acronyms, and refined notation.


References

[1] Panei et al. HARIBOSS: a curated database of RNA-small molecules structures to aid rational drug design. Bioinformatics, 38:4185–4193, 2022.

[2] Stelzer et al. Discovery of selective bioactive small molecules by targeting an RNA dynamic ensemble. Nat. Chem. Biol., 7:553–559, 2011.

[3] Guan et al. 3D equivariant diffusion for target-aware molecule generation and affinity prediction. ICLR, 2023.

[4] Schneuing et al. Structure-based drug design with equivariant diffusion models. Nat. Comput. Sci., 4:899–909, 2024.

[5] Skreta et al. The superposition of diffusion models using the Itô density estimator. ICLR, 2025.

[6] Skreta et al. Feynman-Kac correctors in diffusion: annealing, guidance, and product of experts. ICML, 2025.

[7] Stefaniak and Bujnicki. AnnapuRNA: a scoring function for predicting RNA-small molecule binding poses. PLOS Comput. Biol., 17:e1008309, 2021.

[8] Ontiveros-Palacios et al. Rfam 15: RNA families database in 2025. Nucleic Acids Res., 53:D258–D267, 2025.

[9] Nawrocki and Eddy. Infernal 1.1: 100-fold faster RNA homology searches. Bioinformatics, 29:2933–2935, 2013.

[10] Philips et al. LigandRNA: computational predictor of RNA–ligand interactions. RNA, 19:1605–1616, 2013.

[11] Zhang et al. US-align: universal structure alignments of proteins, nucleic acids, and macromolecular complexes. Nat. Methods, 19:1109–1115, 2022.

[12] Jiang et al. SMARTFlexDB: a paired apo–holo RNA structure database and analysis platform for RNA conformational flexibility and small-molecule recognition. Nucleic Acids Res., gkag1029, 2026.

AuthorsOpenReview ↗

Revision Summary

We highlight all manuscript additions in blue across the 10-page main text and Appendices A–D, indexed below by Section, Equation, Figure, and Table:


1. Notation, Acronym Expansion & Typography — Reviewer LoFT Q8

  • Abstract, Secs. 1–5, Figs. 1 & 3, Appendices A–D: We expanded every acronym at its first appearance, including SE(3)\mathrm{SE}(3), PDB/RCSB PDB, SMILES, SDF, KL, GVP, MLP, SiLU, ESS, NFE, CDF, QED, SA, JSD, AnnapuRNA IMP, OOD, RMSD, MAE, RMSE, SD, SEM, CI, TMpocket\mathrm{TM}_{\mathrm{pocket}}, and vdW radius.
  • Sec. 3.1 & App. C.3–D: We separated atomic features aja_j from residue identities sjs_j in C={(rj,aj,sj)}j=1M\mathcal C = \lbrace (r_j, a_j, s_j) \rbrace_{j=1}^M, wrote Novelty explicitly in Eq. 29, numbered Figs. 12–18, and split wide appendix tables into multi-panel layouts.

2. Controlled Ablations, Backbone Transfer & Apo–Holo Diagnostics — Reviewer 79NT Q3; Reviewer LoFT Q1–Q2; Reviewer oPU9 Q1–Q2, Q4

  • Secs. 1, 3.2–4.3, Table 1, App. A.1 Eqs. 13–14, B.4, D.2 Tables 18–20: We highlighted pψ(N∣C)p_\psi(N\mid\mathcal C) in Sec. 1, added Valid% and five controlled module and count-prior rows to Table 1, documented two-phase rotation at τ≥500\tau \ge 500 and projection/resampling clips, and added Table 18 on the 16-row bridge, DiffSBDD transfer, and 125-PDB / 14-group cluster CIs, Table 19 on synthetic recovery, rotation schedules, and the 186 → 29 → 14 SMARTFlexDB Apo–Holo evaluation, and Table 20 on Fsp3\mathrm{Fsp}^3, SA/Diversity, per-atom proxy LE vs. −8.62 native, and size-matched IMP%.

3. Pretraining Manifest & Single-GPU Runtime — Reviewer LoFT Q6–Q7

  • Secs. 3.1, 4.1–4.2, App. B.2 Tables 4–5 & B.6: We documented the shared CrossDocked2020 v1.1 manifest covering 100,000 training pairs across 15,153 source PDBs and 100 disjoint validation pockets, the 4-stage curation pipeline, and single-A100 parallel 50-particle runtimes at 14.2 s and 313 NFE per particle.

4. Funnel Coverage, Grouped Means, Top-5 Guide & Lead Ranking — Reviewer 79NT Q1; Reviewer LoFT Q3–Q5, Q7; Reviewer oPU9 Q3

  • Secs. 4.1–4.2, Table 2a, App. C.3 Eqs. 24–30 & Tables 13–17: We added stagewise funnel counts confirming 100% scoring coverage and <5%/0%-clash yields in Table 13, 15-model Equal-Pocket/PDB/Group summaries and cluster CIs against TD P→R, TD P, and DiffSBDD P in Tables 14–15, the Top-5 metric guide in Table 16, and Eq. 30 batch lead selection under default <5% and strict 0%-clash riboswitch gates plus 589-pocket LigandRNA validation in Table 17.

5. Neighboring Proteins, OOD Strata & Verification Scope — Reviewer 79NT Q2, Q4; Reviewer oPU9 Q3

  • Secs. 4.1–4.2, 5, Table 2b, App. C.1 Table 9, C.2 Table 12, D.5 Tables 21–24: We quantified 5 Å protein proximity in Table 9, full-assembly RNA and RNA+Protein Clash% in Table 21, the 250-pocket protein-free benchmark in Table 22, six-axis split overlap in Table 12, separately evaluated OOD subsets covering s<0.4s < 0.4, unseen scaffold, unseen Rfam, and TMpocket<0.50\mathrm{TM}_{\mathrm{pocket}} < 0.50 plus all five RNA classes in Tables 23–24, and in silico vs. wet-lab verification scope in Sec. 5.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.