acceptodds
Under review as a conference paper at ICLR 2027

Recursive Constraint Compilation for Guided Protein Backbone Design

Abstract

Many protein design problems require designing a flexible backbone region while leaving the rest of the protein structurally and functionally intact. This is especially important in tasks such as antibody-loop design, fragment inpainting, and motif-linker design, where a useful design must also connect exactly to fixed structural boundaries. Existing methods for constrained generation can enforce hard geometric constraints but typically do so by modifying the generative process itself or by repairing the structures they have generated. To address this gap, we introduce Recursive Constraint Compilation (RCC), a generator-independent framework that recursively decomposes a global hard constraint into local compatibility conditions and assembles exact local parameterizations into explicit coordinates for valid structures. For fixed-endpoint protein backbones, RCC provides exact, invertible coordinates on regular regions of the constrained space, enabling probability modeling, generation, and guidance while maintaining the geometric constraints throughout inference. We evaluate RCC across pretrained protein generators for antibody-loop design, fragment inpainting, and motif-linker design, as well as in prospective CDR-H3 design. Across these settings, RCC preserves the behavior of existing generators while enforcing exact geometry and enables effective guidance over valid backbones that is not recovered by unconstrained Cartesian optimization followed by post-hoc repair.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
constrained generative modeling, protein design, protein backbone generation, geometric constraints, protein structure generation

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_oaPnOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3(+1)
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes RCC, a generator-independent framework that enforces hard geometric constraints by compiling them directly into the representation space, rather than relying on penalty losses, modified generative dynamics, or post-hoc projection. By recursively decomposing global constraints into local compatibility conditions, RCC constructs exact, invertible, and differentiable coordinates for valid protein backbones, enabling constraint-preserving generation, probabilistic modeling, and inference-time guidance without retraining pretrained generators. Experiments demonstrate geometric feasibility and structural quality comparable to conventional projection methods. In CDR-H3 design, RCC-coordinate guidance improves structural success from 42.0% to 70.1%, substantially outperforming unconstrained Cartesian optimization followed by post-hoc repair (6.6–11.1%).

These results highlight the value of encoding hard constraints directly into the representation, allowing generation and optimization to remain within the feasible space throughout inference.

Critical strengths

I believe that a key strength of this paper lies in its conceptually interesting and mathematically grounded reformulation of constrained protein generation.

Rather than enforcing hard constraints through penalty losses, modified generative dynamics, or post-hoc corrections, RCC constructs a coordinate representation of the feasible space, effectively transforming a challenging constrained optimization problem into a more tractable optimization problem over coordinates that preserve feasibility by construction. I find this to be an elegant and insightful approach to handling hard constraints in probabilistic generative modeling.

Moreover, the method is supported by detailed mathematical derivations and theoretical results concerning coordinate invertibility and volume correction, along with empirical improvements over several relevant baselines.

Critical weaknesses

Although RCC demonstrates substantial improvements over unconstrained optimization followed by post-hoc repair, its performance is comparable to other feasibility-preserving baselines, particularly Riemannian guidance.

This makes it difficult to determine whether the observed improvements stem primarily from maintaining feasibility throughout optimization rather than from the specific advantages of the proposed coordinate representation. While this does not diminish the theoretical contribution of RCC, it somewhat limits the empirical evidence for its distinct practical advantages.

Questions for authors

Thank you to the authors for this interesting work. I have two questions that might help the authors improve:

  • Could the authors further clarify the distinctive practical advantages of the proposed coordinate representation compared with existing feasibility-preserving methods, particularly given their comparable performance in Appendix E.5?
  • Given that the theoretical construction requires recursively composable constraints with bounded-dimensional summaries, could the authors clarify what evidence supports the broader applicability of RCC beyond the two demonstrated constraint families?
3 comments

We thank the reviewer for recognizing RCC's mathematically grounded formulation. To clarify the advantages of its coordinate representation, we compared stronger feasibility-preserving methods and evaluated a task that uses RCC's probabilistic structure directly.

W1 + Q1. What does RCC offer beyond feasibility preservation?

This makes it difficult to determine whether the observed improvements stem primarily from maintaining feasibility throughout optimization rather than from the specific advantages of the proposed coordinate representation.

We added projected Adam + Newton and expanded the Riemannian + Newton comparison under matched conditions. On the original 288-condition panel (255 PDBs), methods used three checkpoints and three search seeds, shared starts and objectives, 15 updates, and 64 scored calls.

MethodCombined successRCC − method, pp (95% CI)
RCC-coordinate Adam48.6%Reference
Riemannian gradient + Newton50.7%−2.1 [−4.8, +0.7]
Projected Adam + Newton48.8%−0.3 [−3.1, +2.6]
Adam differentiated through Newton45.7%+2.9 [+0.0, +5.7]

Combined success requires geometric success and PULCHRA steric cleanliness. Intervals use paired PDB-cluster bootstraps, and this expanded test-panel study is post hoc. The results show RCC performs effectively with strong feasibility-preserving guidance methods. Matched RCC, NAMM + Newton, and diffusion + Newton generators also achieved 87/96, 88/96, and 80/96 combined successes after guidance and screening on reused development PDBs.

RCC's distinctive practical benefit is more accurate reconstruction of missing protein loops from sparse, noisy structural measurements. The explicit inverse and tractable density enable local probabilistic inference within the feasible backbone space. We tested this on 229 held-out complexes, with eight simulated noisy distances and matched prior, context, starting states, Cα constraints, and 30-second CPU budgets.

Reconstruction methodMean Cα RMSD ↓Within 2 Å
RCC density-corrected inference2.869 Å133/458
Newton inference3.094 Å105/458
Density-free suffix inference3.119 Å111/458
Direct sampling3.151 Å96/458

RCC achieves the lowest mean error and highest observed recovery count. Its mean-error reductions are 7.3% versus Newton (adjusted 95% CI [4.3%, 9.1%]), 8.0% versus suffix inference [4.7%, 10.3%], and 9.0% versus direct sampling [5.0%, 11.2%]. All methods share the learned prior and geometric constraints, isolating the benefit of RCC's density-corrected inference procedure. The intrinsic-volume term is not independently isolated.

The reconstruction benchmark provides direct empirical evidence of the value of RCC's probabilistic coordinates beyond feasible guidance. Optimizer and generator controls appear in Section 3.4 and Appendices H.2–H.3 and H.16–H.17. The reconstruction benchmark will be incorporated into Section 3.6 and Appendix H.15.

Q2. Broader applicability of recursively composable constraints

Given that the theoretical construction requires recursively composable constraints with bounded-dimensional summaries, could the authors clarify what evidence supports the broader applicability of RCC beyond the two demonstrated constraint families?

We developed two additional constructions beyond fixed-endpoint Cα bridges and bounded fixed-total allocations, while preserving the requirement for bounded-dimensional recursive summaries.

Richer peptide backbone. An N/Cα/C/O chart combines RCC's bridge coordinates with rotations of rigid planar peptide units. It enforces specified bond lengths, peptide-unit angles, trans planarity, and fixed-anchor attachment while retaining explicit encoding and decoding. The N–Cα–C angle remains softly restrained. Maximum numerical errors are 2.5 × 10⁻¹⁴ Å for specified bonds, 1.8 × 10⁻¹² degrees for peptide-unit angles and planarity, and 2.9 × 10⁻¹¹ Å for encode–decode reconstruction. Native H3 backbones have a median all-atom representation error of 0.08 Å.

On 202 generated Context-FrameChart development backbones, the new representation improved geometric quality.

MeasureBeforeAfter
Bond outliers per 100 checks20.30.0
Angle outliers per 100 checks47.22.0
Ramachandran-favored residues53.6%75.9%

RCC extends its exact, invertible representation to chemically meaningful peptide geometry, eliminating specified bond outliers and improving backbone-angle and Ramachandran statistics. With identical peptide constraints, RCC and Newton reached metric-preconditioned objectives of 2.499 and 2.496, respectively.

Certified full-atom construction. A separate extension uses six-dimensional relative-pose summaries and admits a chart only when its validity checks pass. On 20 new CATH loops, 16 admitted valid charts and 14 met the prescribed 0.1 Å motion target. All 20 underwent independent CCTBX/Probe validation. We retain unsuccessful admissions in the denominator.

The peptide and rigid-pose constructions provide concrete evidence that RCC extends to richer protein constraints while retaining exact, invertible coordinates. The theory specifies the applicable class through bounded-dimensional recursive summaries and valid local charts. Side-chain packing and energetics can be modeled separately. The representation appears in Section 2.3, with validation and optimizer controls in Appendices H.4–H.5 and H.12–H.13.

We thank the reviewer again for the thoughtful assessment. We hope the expanded feasible comparisons, practical reconstruction result, and richer-constraint studies address the remaining questions. In light of this evidence, we would be grateful if the reviewer would consider increasing their score and continuing their support of our submission. Thank you!

Thank the authors for their thoughtful responses and the additional experiments. I believe my concerns have been well addressed.

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

The paper focuses on a problem of repairing protein structures to enforce structural constraints. This problem has many applications in biology, like antibody-loop design, fragment inpainting and motif scaffolding, which requires to design a flexible backbone region while leaving the rest of the protein structurally and functionally intact. To address the problem, the authors introduce Recursive Constraint Compilation, which recursively decomposes a global constraint into local compatibility conditions. The authors mainly focus on the constraint of fixed endpoints and Cα–Cα distances in this paper. They performed experiments to demonstrate RCC preserves the behavior of existing protein generators while enforcing exact geometry.

Critical strengths
  1. The paper presents a clear geometric method to connecting a backbone between two fixed endpoints while keeping Cα spacing constant. The mathematical derivations and numerical in Sec. 3 checks support that it works as intended.
Critical weaknesses
  1. It is unclear whether the method produces useful protein designs. Better Cα geometry does not consistently improve atom-level geometry, and none of the 64 sequences tested for the main guided method recovered the designed loop within 2 Å after sequence re-design in Table 26. As protein sequences are delivered to wet-lab for design in practice, it is hard to say the proposed method improves the generated proteins, given it only outputs structures that are incompatible with designed sequences.
  2. The method enforces only fixed endpoints and Cα–Cα distances. These constraints ensure chain connectivity but do not guarantee realistic protein geometry. Important factors, including backbone angles and dihedrals, side-chain conformations, steric clashes, and energetics, are not enforced by the RCC construction itself. Although some can be included in the guidance objective, the current results leave unclear whether satisfying this limited set of hard constraints translates into better protein designs.
Questions for authors
  1. What exactly counts as “structural success” in Table 3? Could you provide the full objective and success thresholds? This would help clarify what the reported 70.14% success rate means for protein design.
  2. The motif-linker experiment uses FrameFlow. Would a stronger, more recent generator, e.g. La-Proteina, already produce sufficiently accurate geometry without repair? It would be helpful to test how often repair is needed and whether it improves downstream design quality with a stronger generator.
  3. The proposed method makes endpoints and Cα spacing exact, but realistic proteins also depend on backbone angles, dihedrals, side-chain packing, and energetics. How much practical benefit comes from enforcing these two constraints alone, and could the framework incorporate richer structural constraints?
  4. In which protein-design applications are RCC’s invertibility and density calculation useful? Since the guidance experiments do not use the density calculation, could you show a practical benefit that depends on this part of the method?
5 comments

We thank the reviewer for recognizing RCC's mathematical formulation. We clarified what the original geometric-success metric measures and expanded the downstream refolding evaluation.

Q1. Definition of structural success

What exactly counts as “structural success” in Table 3? Could you provide the full objective and success thresholds?

The 70.14% in Table 3 is a geometric design success rate, not a measure of folding, binding, or native-loop recovery. A successful candidate satisfies all four criteria.

CriterionThreshold
Separation from epitope, framework, and nonlocal loop Cα atoms≥2.0 Å
Declared epitope contactsAt least min(3, L) loop Cα atoms within 6 Å
Cα(i)–Cα(i+2) distances outside [5.0, 7.4] ÅAt most 25%
Link and endpoint residuals≤5 × 10⁻⁵ normalized units

The original guidance objective includes epitope contacts (weight 0.5), clash/local-geometry penalties, and a base-noise tether (0.01). The new matched feasible-optimizer comparisons use internal repulsion weight 0.25 and no tether. We distinguish the optimization objective from evaluation thresholds and additionally report combined success, which requires PULCHRA steric cleanliness.

W1. Does RCC generate refoldable designs?

None of the 64 sequences tested for the main guided method recovered the designed loop within 2 Å after sequence re-design in Table 26.

To address the original 0/64 guided-refolding result, we evaluated a stronger RCC-coordinate generator, Context-FrameChart, which also models residue frames. On 255 PDBs, we generated two backbones per PDB, designed two ProteinMPNN sequences per backbone, and evaluated framework-aligned H3 Cα self-consistency ≤2 Å using two independent predictors.

BackboneABodyBuilder3 refoldsIgFold refolds
RCC-native generation106/1,020 (10.4%)11.4%
Native-backbone redesign70/510 (13.7%)12.5%

RCC-native refolding rates are close to the matched native-backbone redesign reference, with differences of −3.3 pp [−7.0, +0.3] for ABodyBuilder3 and −1.2 pp [−4.9, +2.4] for IgFold. With the same generator starts, RCC guidance increases geometric success from 39.6% to 77.6% and yields 60/1,020 refolds. We report geometry and refolding separately because the guidance objective optimizes antigen contacts, not predicted sequence recovery.

RCC coordinates support refoldable H3 generation, with 106 successful ABodyBuilder3 predictions and corroborating results from IgFold. This provides the downstream self-consistency evidence missing from the original Cα-only evaluation. We use a stronger generator that also models residue frames, and the results measure computational refolding rather than experimental folding or binding. The success criteria appear in Section 3.3 and the expanded study in Section 3.5, Table 8 and Appendix H.14.

W2 + Q3. Extending RCC to more realistic backbone constraints

The method enforces only fixed endpoints and Cα–Cα distances. These constraints ensure chain connectivity but do not guarantee realistic protein geometry.

To address this concern, we developed an N/Cα/C/O construction combining RCC bridge coordinates with rotations of rigid planar peptide units. The new chart enforces specified bond lengths, peptide-unit angles, trans planarity, and fixed-anchor attachment by construction while retaining explicit encoding and decoding. The N–Cα–C angle is softly restrained rather than exactly fixed.

Maximum numerical errors are 2.5 × 10⁻¹⁴ Å for specified bonds, 1.8 × 10⁻¹² degrees for peptide-unit angles and planarity, and 2.9 × 10⁻¹¹ Å for encode–decode reconstruction. The median native-H3 all-atom representation error is 0.08 Å, with a 90th percentile of 0.51 Å.

To test whether the extension improves generated structures, we fitted 202 Context-FrameChart development backbones with angle and Ramachandran restraints.

Geometric measureBeforeAfter
Bond outliers per 100 checks20.30.0
Angle outliers per 100 checks47.22.0
Ramachandran-favored residues53.6%75.9%

RCC eliminates the specified bond-length outliers and substantially improves backbone geometry, with only 0.93 Å median coordinate displacement. We separately evaluated sequence-dependent refolding, with details in Appendix H.13. The representation also supports effective constrained optimization, reaching a metric-preconditioned objective of 2.499 versus 2.496 for Newton under identical peptide constraints. A certified full-atom construction using six-dimensional relative-pose summaries admitted valid charts on 16/20 new CATH loops and reached the motion target on 14/20.

Together, these experiments demonstrate that RCC's exact, invertible construction extends to chemically meaningful peptide geometry and certified local full-atom constraints. These geometric guarantees complement separate models of side-chain packing and energetics. The construction and validation appear in Section 2.3 and Appendices H.4–H.5 and H.12–H.13.

Q2. Exact geometric correction with La-Proteina

Would a stronger, more recent generator, e.g. La-Proteina, already produce sufficiently accurate geometry without repair? It would be helpful to test how often repair is needed and whether it improves downstream design quality.

We evaluated La-Proteina on 192 motif scaffolds across six tasks. It placed motifs accurately, with median backbone RMSD 0.10 Å, but all 192 structures had at least one peptide C–N bond error greater than 0.05 Å. The median worst C–N error was 0.13 Å. RCC repaired the prescribed geometry with a median coordinate displacement of just 0.07 Å.

For each corrected or uncorrected backbone, we designed eight ProteinMPNN sequences with fixed motifs and evaluated ESMFold self-consistency.

OutcomeLa-ProteinaAfter RCC
Motif-scaffolding success62.0%63.0%
Designable, self-consistency RMSD ≤2 Å72.4%71.4%

Motif success requires self-consistency RMSD ≤2 Å and motif RMSD ≤1 Å. The paired success difference is +1.0 pp [−2.6, +4.7]. RCC enforces the prescribed peptide geometry while preserving La-Proteina's motif accuracy and broadly similar downstream designability.

We also tested La-Proteina on antibody H3 loops attached to fixed frameworks. Raw samples had a median graft-junction peptide error of 0.49 Å. Exact RCC closure preserved geometric success (24%) and changed observed refolds from 15/256 to 17/256. Antigen-specific RCC guidance yielded 22/256 refolds and 10/256 joint geometric-and-refolding successes, versus 16/256 and 8/256 with Riemannian + Newton. These experiments show that RCC can both correct graft geometry and support antigen-specific guidance with a strong pretrained generator.

The practical benefit is that RCC can correct geometric violations in a modern pretrained generator without retraining it or substantially perturbing its structures. The La-Proteina experiments, including computational designability and refolding outcomes, appear in Appendix H.8.

Reviewer_ie9vOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Recursive Constraint Compilation (RCC), a decomposition approach for global constraints imposed over protein backbone atomic coordinates. The approach operates by recursively splitting the chain into subchains, resolving these independently and parameterizing the meeting point of these subchains. The approach is presented as a general method, independent of a specific generation process, with existing generators being paired with RCC for the empirical validation. The paper headlines improved success rates when applying RCC in the experiments.

Critical strengths
  • Generality of the Framework: The proposed contribution is not tied to a specific generator, which makes the approach generally applicable. This is appealing, supporting its potential adoption in various protein design settings.
  • Feasible-by-Design Formulation: Endpoint and link-length constraints are feasible by construction via De\mathcal{D}_{e}, which supports the utility of the proposed system.
Critical weaknesses

Experimental Validation: The experiments are difficult to interpret considering several missing pieces.

  • Baselines are limited. The main comparison (Table 3) motivates the use of constraints; however, the table doesn’t capture whether other constrained methods would achieve similar performance.
  • Other feasibility maintaining methods perform similar. For example, “Riemannian gradient + Newton” outperforms RCC in Table 25 and is statistically indistinguishable in Table 24. I also think this comparison should be highlighted in the main text, since it is more interesting than RCC compared to constraint-agnostic approaches.

Downstream Results: Table 26 shows that none of the designs pass the self-consistency threshold. This makes the upstream metrics difficult to trust, as they are not a strong indicator of downstream utility.

Extension to General Constraints: It remains to be established whether this method can extend to more general constraint classes that are present in broader protein design settings, with RCC being specifically adept at particular endpoint constraints but perhaps not general to settings.

Questions for authors
  • Could the authors please provide more details on the pertained generators used in the evaluation? I couldn’t find many details on the specific model, architecture, etc.
6 comments

We thank the reviewer for recognizing RCC's generator-independent, feasible-by-construction formulation. We expanded the feasibility-preserving baselines and tested a further practical use of its invertible probabilistic representation.

W1 + W2. Comparison with feasibility-preserving methods

Other feasibility maintaining methods perform similar. For example, “Riemannian gradient + Newton” outperforms RCC in Table 25 and is statistically indistinguishable in Table 24.

We agree that feasibility-preserving methods provide the most informative comparison, and we now highlight these results in the main text. RCC optimizes invertible coordinates that decode to feasible backbones, projected Adam takes Cartesian updates followed by Newton closure, and Riemannian guidance takes tangent updates with Newton closure. All enforce the same Cα constraints.

On the original 288-condition panel (255 PDBs), methods share three proposal checkpoints, three search seeds, four starts, the same contact/clash objective, 15 updates, 64 scored calls, and best-objective selection.

MethodCombined successRCC − method, pp (95% CI)
RCC-coordinate Adam48.6%Reference
Riemannian gradient + Newton50.7%−2.1 [−4.8, +0.7]
Projected Adam + Newton48.8%−0.3 [−3.1, +2.6]
Adam differentiated through Newton45.7%+2.9 [+0.0, +5.7]

Combined success requires Cα structural success and PULCHRA steric cleanliness. Intervals use paired PDB-cluster bootstraps. Step sizes were selected on development data, and the expanded test-panel comparison is post hoc. Further feasible controls, including FABRIK and RCC retraction, appear in Appendix H.2.

The optimizer comparisons establish RCC as an effective method for feasible guidance. Its distinctive practical benefit is more accurate reconstruction of missing protein loops from sparse, noisy structural measurements. RCC's explicit inverse and tractable density enable local inference over the same constrained backbone space. We tested this on 229 held-out antibody–antigen complexes, using eight simulated noisy distances. Every method shared the learned prior, protein context, 32 starting states, Cα constraints, and a 30-second CPU budget.

Reconstruction methodMean Cα RMSD ↓Recovered within 2 Å
RCC density-corrected inference2.869 Å133/458
Newton inference3.094 Å105/458
Density-free suffix inference3.119 Å111/458
Direct sampling3.151 Å96/458

RCC achieves the lowest reconstruction error and highest observed recovery count among all four methods. It reduces mean error by 7.3% versus Newton (adjusted 95% CI [4.3%, 9.1%]), 8.0% versus suffix inference [4.7%, 10.3%], and 9.0% versus direct sampling [5.0%, 11.2%]. Since all methods satisfy the same geometric constraints, this advantage goes beyond feasibility preservation and illustrates the benefit of RCC's probabilistic inference procedure.

Together, the two experiments show RCC's value in both constrained guidance and recovering missing protein geometry from incomplete observations. We report the optimizer controls in Section 3.4, Table 5 and Appendices H.2–H.3. The full reconstruction benchmark will be incorporated into Section 3.6 and Appendix H.15.

W1. Comparison with constrained generative methods

Baselines are limited. The main comparison (Table 3) motivates the use of constraints; however, the table doesn’t capture whether other constrained methods would achieve similar performance.

We added matched constrained generative baselines. Autoregressive RCC, an H3 adaptation of Neural Approximate Mirror Maps (NAMM), and diffusion + Newton were trained on the same 1,433 loops, with three seeds and matched model-level training budgets. We evaluated them on 32 reused development PDBs, with the same candidate screen and guidance objective.

GeneratorScreen onlyGuidance + screen
Autoregressive RCC55/9687/96
NAMM H3 adaptation + Newton68/9688/96
Diffusion + Newton53/9680/96

These are combined geometric-and-steric successes over 32 PDBs and three seeds. With four starts and 32 updates, RCC improves by 33.3 percentage points [21.9, 45.8] and performs comparably to NAMM after guidance (RCC − NAMM, −1.0 pp [−6.3, +4.2]). The same-start guidance controls, selection stages, training details, and crossed-seed uncertainty appear in Appendices H.16–H.17, including the implementation details of our H3-specific NAMM adaptation.

Correction inside one pretrained generator. To isolate the repair mechanism, we also compared RCC trajectory correction with an H3 adaptation of the proximal method of Christopher et al. (2026) using the same RFdiffusion Base checkpoint, 32 development PDBs, four seeds, and 50 denoiser evaluations.

Correction methodFeasibleCombined successMean generation time
RCC trajectory correction128/12826/12823.7 s
Christopher proximal correction, H3 adaptation123/12825/12872.9 s

RCC repairs all 128 outputs, with 26/128 combined successes versus 25/128 for the proximal adaptation (+0.8 pp [−2.3, +3.9]). It also requires 23.7 versus 72.9 seconds, about one-third the runtime. The timing comparison applies to this H3 adaptation of the correction mechanism, with full protocols and failure accounting in Appendix H.9.

RCC can enforce exact geometry in an existing pretrained protein generator without retraining it. Alongside the matched generator comparisons, this provides a practical demonstration of RCC as a reusable tool for constrained protein generation and guidance. We report the studies in Section 3.4, Tables 6–7.

W3. Downstream refoldability of RCC-native generation

Table 26 shows that none of the designs pass the self-consistency threshold. This makes the upstream metrics difficult to trust, as they are not a strong indicator of downstream utility.

The original 0/64 result came from the initial Cα-based guided proposal. To test whether RCC coordinates can support refoldable designs with richer structural modeling, we evaluated Context-FrameChart, which also models residue frames. Across 255 PDBs, we generated two backbones per PDB, designed two H3 sequences per backbone with ProteinMPNN, and evaluated framework-aligned H3 Cα self-consistency (≤2 Å) using ABodyBuilder3 and IgFold.

BackboneABodyBuilder3 refoldsIgFold refolds
RCC-native generation106/1,020 (10.4%)11.4%
Native-backbone redesign70/510 (13.7%)12.5%

RCC-native generation achieves refolding rates within a few percentage points of native-backbone redesign under the same pipeline. The paired differences are −3.3 pp [−7.0, +0.3] with ABodyBuilder3 and −1.2 pp [−4.9, +2.4] with IgFold.

RCC guidance from the same starting backbones increased geometric success from 39.6% to 77.6% and yielded 60/1,020 ABodyBuilder3 refolds. Because this guidance optimizes contact geometry rather than predicted refolding, we report these outcomes separately, with matched Riemannian + Newton results in Appendix H.14.

RCC-coordinate generation produces self-consistent H3 backbones under both structure predictors, at rates approaching native-backbone redesign with the same evaluation pipeline. This directly addresses the original concern about downstream structural plausibility. The study uses a stronger model that also generates residue frames and evaluates computational refolding, with experimental validation remaining a separate question. We report it in Section 3.5, Table 8 and Appendix H.14.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.