acceptodds
Withdrawn from ICLR 2027

Co-DIRECT: Adaptive Coordination of Sequence and Structure for Protein Design

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Joint generative models aim to produce compatible outputs by allowing their predictions to inform one another. In protein design, this requires coordinating an amino-acid sequence with its three-dimensional structure. Yet stronger influence between these predictions does not necessarily make that influence beneficial. We investigate this distinction in a joint protein generator and find that the benefit of an additional prediction varies across residues and tasks, while the resulting response can increase reference-based prediction error. We introduce Co-DIRECT to adaptively coordinate sequence and structure through two components: anticipation and a router. Anticipation probes how the model's predictions respond to a tentative update, providing corrections that account for the coupled response. The router uses information from both sequence and structure representations to determine which representation receives these corrections at each residue. Together, these components regulate generation without querying an external structure predictor during sampling. Across binder design, enzyme scaffolding, and antibody CDR redesign, Co-DIRECT improves the reported mean assessments of refolded designs relative to frozen and fully adapted Proteina-Complexa. Component ablations and routing controls support the combined approach, with improved outcomes accompanied by fewer local error-increasing responses. These findings support adaptive coordination of sequence and structure as a strategy for improving joint protein generation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein design, Joint sequence-structure generation, Flow matching, Multimodal generative models, Adaptive routing

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_6aViOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces Co-DIRECT, a method that improves joint protein sequence–structure generation by actively controlling how the two modalities, backbone coordinates and sequence latents, influence each other during sampling. The approach combines anticipation with a residue-wise directional router to determine which channel leads and which follows. Benchmark results show that Co-DIRECT reduces local errors and improves benchmark metrics compared with Proteina-Complexa and the RFdiffusion+MPNN pipeline.

Critical strengths
  1. Residue-wise routing serves as a control mechanism across the sequence and structure modalities, guiding coordination between the two modalities in terms of their rates of progression and quality.
  2. The work highlights that enabling cross-modal interactions does not necessarily ensure beneficial guidance. Turning these interactions into an explicit control mechanism may therefore be a meaningful step.
Critical weaknesses
  1. The ablation experiments remain incomplete. In the inference-time evaluations of fixed and shuffled roles, the reported intervals often include 0, which makes the residue-wise router appear unnecessary. Furthermore, the “no router” ablation changes several settings simultaneously, including roles and self-conditioning weights, making it difficult to isolate and clearly interpret the effects of individual components.
  2. The evaluation omits many co-design baselines, such as PXDesign. The limited and relatively dated selection of baselines makes it difficult to assess the value of the work accurately.
  3. The presentation lacks clarity. Many important numerical results are embedded in the text rather than presented systematically in tables or figures.
Questions for authors
  1. Could the authors provide more informative ablation experiments, including ablations of the router’s individual components, and clarify why residue-wise routing is necessary?
  2. Do the components of Co-DIRECT introduce substantial computational overhead? Given that downstream tasks often require a large number of design inference runs, can Co-DIRECT produce strong results with low computational overhead?
Reviewer_YVEYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Co-DIRECT adapts a pretrained joint protein sequence-structure flow model using two generator calls per sampling step. At each eligible designed residue, a learned router assigns leading and responding roles to the coordinate and residue-latent channels. Anticipation probes the leader through a geometry snap to a nearby partner atom or a short, clipped latent lookahead. The update combines the first-call leader velocity with the second-call follower velocity and integrates from the original state. Router scores weight self-conditioning in both calls. The second call recomputes these weights but retains the original roles. An auxiliary loss supervises the follower response against the original-state target. The paper also introduces a paired diagnostic that separates decreases and increases in prediction error for each residue and channel under a fixed scoring state and reference. Evaluation covers 30 binder, 41 enzyme, and 32 antibody targets, with three candidates per target and five refolds per candidate under common refolding settings. Relative to ordinary adaptation of the same generator, mean binder ipTM increases from 0.309 to 0.547 and mean enzyme ligand-fit RMSD falls from 19.89 to 8.16 Å.

Critical strengths
  1. The gains over ordinary adaptation are large and persist when the number of sampling calls is approximately matched. Mean binder ipTM rises from 0.309 to 0.547 and enzyme ligand-fit RMSD falls from 19.89 to 8.16 Å, with 26 of 30 binder targets and 39 of 40 paired enzyme targets improving (Tables 17-19). Within each task family, Full and ordinary adaptation use the same initial checkpoint, training data, molecular conditions, and number of optimizer updates (Appendix A.8). Increasing ordinary sampling from 400 to 799 calls improves binder ipTM by only 0.0088 and enzyme RMSD by 0.49 Å (Figure 4a).

  2. The evaluation measures structural consistency as well as confidence under a common protocol. The protocol allocates three candidates per target to each method, with five ESMFold2-Fast refolds per candidate under shared settings. A common rule within each task family selects one candidate and refold for all endpoints, avoiding separate selection of the best result for each metric (Appendix B.6). Design comparisons bootstrap paired targets, while the Figure 2 diagnostic bootstraps source proteins, so repeated measurements are not treated as independent observations (Appendices B.4 and B.9). Mean Cα scRMSD between the generated structures and their selected refolds falls from 16.90 to 4.61 Å for binders and from 18.02 to 7.42 Å for enzymes. These coordinate comparisons support improved sequence-backbone agreement. Antibody ipAE also improves from 19.10 to 13.32 Å, although the 95% confidence interval for the primary ipTM difference includes zero (Tables 17, 21-22, 25).

Critical weaknesses
  1. The ablations do not establish the separate benefits of learned role assignment and response-loss placement. On binders, forcing structure to lead reduces ipTM by 0.0144, forcing sequence to lead by 0.0173, and shuffling roles among residues at comparable partner distances by 0.0483. Across all three tasks, the primary-endpoint intervals for these role interventions include zero, as do those for the retrained fixed-direction control (Tables 18-20). The larger gap for the no-router control does not isolate role assignment. This variant also changes the probe to a symmetric Euler probe and restores ordinary self-conditioning (Appendix D.2). Removing anticipation reduces performance, supporting the combined query and response objective, but leaving the benefit of adaptive roles unresolved. Moving the auxiliary objective from the second prediction to the first reduces binder ipTM by 0.0055, increases enzyme RMSD by 0.148 Å, and reduces antibody ipTM by 0.0092. All three intervals include zero. The primary design endpoints do not show an advantage for supervising the second-call response over applying the extra objective to the first prediction.

  2. The enzyme results show lower average error, but none of the selected designs meets the stated ligand-fit criterion. All 41 Co-DIRECT targets exceed the 3 Å threshold. The minimum is 4.62 Å, the median 7.33 Å, and the mean 8.16 Å (Appendix E.1 and Table 24). Ligand-fit RMSD scores the catalytic motif atoms after aligning the ligand to its reference, retaining errors in both motif geometry and placement relative to the ligand. The mean motif-fit RMSD of 1.81 Å does not establish correct placement, because fitting the motif itself removes its overall rigid displacement. The authors report the zero pass rate, and RFdiffusion2 also has no passing targets under this protocol. The paper does not identify the stage at which this common failure occurs. Comparing catalytic geometry in the generated and refolded structures would help identify whether the errors are already present at generation or arise or worsen during refolding. The reported mean reduction is substantial, but the enzyme evaluation does not yet demonstrate a selected design satisfying the paper’s geometric criterion. This limits the evidence for useful enzyme scaffolding under the reported protocol.

Questions for authors
  1. Can the authors report a retrained fixed-direction control that retains router-weighted self-conditioning, the same directional probe construction rules, and the training objectives used by Full? The existing retrained control uses ordinary self-conditioning (Appendix D.4), while the inference interventions reuse Full and leave the primary-endpoint differences unresolved. The no-router control changes both the probe and conditioning pathway. A matched comparison on one existing task family would test the value of learned residue-wise assignments while retaining the rest of the procedure. Can the authors also explain why second-call supervision is preferred over the original-state auxiliary control, given the unresolved differences in Tables 18-20?

  2. At what stage do the enzyme designs fail the ligand-fit criterion? All 41 selected Co-DIRECT designs exceed 3 Å, with a minimum of 4.62 Å and a median of 7.33 Å (Appendix E.1 and Table 24). The mean motif-fit RMSD of 1.81 Å does not establish correct placement relative to the ligand, since that alignment fits the motif itself. Where catalytic-atom correspondence permits, can the authors compare both RMSDs in the generated structures and their selected refolds? RFdiffusion2 also has no gate passes under this protocol. Examining these existing outputs would help determine whether the geometry errors are present before refolding or arise or worsen during refolding.

Reviewer_x4VWOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The authors of the paper seek to study whether the capabilities of coupled sequence and structure generation are maximized. To this end, they first demonstrate a proof of concept experiment on Proteina-Complexa predicted velocity fields and then they extend this further to show that interactions within a joint generator can be measured and controlled. The authors evaluate their framework over different binder designs, enzyme scaffolding, and antibody CDR-H3 redesigns.

Critical strengths
  1. The authors present the failure modes with complete transparency. On enzyme design, their method fails to meet the defined threshold.

  2. The authors also provide the necessary code to reimplement the method

  3. All methods go through a carefully designed evaluation framework. This includes a shared refold budget and a single selection rule across all endpoints. More importantly, the authors also distinguish between motif-fit and ligand-fit.

Critical weaknesses

Overstated claims

The authors claim that current jointly trained protein design models like Proteina-Complexa are not using the coupling well. While the claim makes sense, the authors do not show that making appropriate usage of the couplings helps gain an improvement on top of what Proteina-Complexa already offer. The paper instead demonstrates on Complexa with its test-time search disabled, at binder ipTM 0.241, which is not a well-functioning joint generator.

iPTM is used as design-quality endpoint

In the paper, the authors primarily evaluate their binder and antibody using iPTM, which is a confidence score that can be easily manipulated without significantly improving designs.

Response loss with zero gradient

Appendix A.4 sets h = 0 on half of training batches, and Appendix A.7 notes the router's gradient from prediction losses vanishes when h = 0. This means, the router only receives geometric supervision. This seems dubious.

Few candidates per target

Typical design benchmarks report success rates over hundreds to thousands of samples. By setting G to 3, the reported means are heavily influenced by sampling luck per target.

Diagnostic in section 3 and method in section 4 measure different things

In section 3, the authors propose a proof of concept experiment which would apply a uniform and symmetric correction term. The experiment primarily controls the strength of the correction term, which can deteriorate the design quality. This supports the need for an attenuation term. However, Co-DIRECT does not do that. Eq 7 replaces the follower's velocity with the second-call value, with no attenuation. The authors do run the experiment in Appendix B.3 to show the attenuation over different values of alpha. But, no values below 1 are reported.

Interface loss ablation missing

Every variant in Table 15 carries the contact/interface loss at coefficient 2. So no experiment can tell whether the improvement comes from the router and anticipation or simply from the interface loss that every variant adds.

Questions for authors
  1. Can the authors compare Co-DIRECT against Complexa with reward-guided search enabled, or at a candidate budget where its reported in-silico success rates are reproduced? If routed anticipation helps there, the central claim makes sense

  2. Do the binders survive a non-confidence based metric, like interface geometry metrics?

  3. How do the paired differences behave as the candidate budget grows?

  4. Can the authors report the deterioration frequency at different values of alpha less than 1?

  5. Can the authors report ordinary full adaptation trained with the interface loss at coefficient 2, without the router or probe? Without it, the gain over adaptation cannot be attributed to anticipation or routing rather than to the added contact supervision.

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

Related papers

Discussion (0)

Sign in to comment.