acceptodds
Under review as a conference paper at ICLR 2027

CODesign: Consistency from Data to Trajectory in All-Atom Protein Binder Co-Design

10 views

Abstract

The central challenge in protein design is generating plausible, mutually compatible structures and sequences, such that each designed sequence folds into its intended structure and the structure accommodates that sequence. Compared to typical two-stage design methods, which decouple the modeling of the interdependent modalities, co-design models improve the cross-modal consistency by jointly generating sequences and structures. However, naively generating sequences and structures simultaneously does not ensure their consistency. To address this challenge, we propose CODesign framework. We improve data consistency by generating approximately 105,000 consistency-distilled dimers. We further promote consistency through a multimodal joint flow model that captures the joint distribution of sequences, backbone structures, and local atomic configurations, together with a consistency-aware joint resampling strategy that iteratively refines sequences and side chains. Experiments show that CODesign achieves state-of-the-art performance with the highest in silico success rates on both protein- and ligand-target binder design. Ablation studies also demonstrate our distilled dataset increases performance by 70.9%, and can be further improved by our proposed resampling mechanism with negligible additional computational cost. Code, model weights and the new dataset will be completely open-source.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
binder design, protein design, generative modeling

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_SaBHOpenReview ↗
Rating
3
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper works on the problem of de novo protein binder design with a model explicitly modeling sequence, structure and side-chain information. This work can be seen as a direct extension of Proteina-Atomistics on the task of binder design. The authors adapt the same architecture with an additional target conditioning module and propose to use consistency-distilled dimers to improve data consistency similarly as Proteina-Atomistica. The authors further propose a consistency-aware joint resampling strategy that iteratively refines sequences and side-chains. The experimental results do show improvements over latent-base methods like Complexa on both protein and ligand binder design.

Critical strengths
  1. Though the idea is similar to Proteina-Atomistica, it is still very interesting to show that explicitly modeling sequences and structures can work well for binder design. The extension of the model from monomer generation to binder generation is non-trivial, as seen in different mechanisms introduced in this paper, like consistency-distilled dataset and consistency-aware joint resampling.
  2. The improvement over Complexa on ligand-conditioned protein binder design is significant, which clearly shows the advantages of the model on in-silico benchmarks.
Critical weaknesses
  1. The main weakness of the paper is the novelty against Proteina-Atomistica. The model architecture and data distillation idea closely follow Proteina-Atomistica. The main methodological contribution is only the consistency-aware joint resampling strategy. That said, this weakness alone shouldn't lead to the rejection of the paper.
  2. Another potential concern is the additional time overhead introduced by atom encoder/decoder. The authors only report timing for ligand-conditioned binder design, which is ~4x slower than Complexa. Though not reported, I assume that for proteins, the model will be even slower, given the larger number of atoms in protein targets. This might limit the performance with inference-time search with a fixed time budget.
Questions for authors
  1. For Figure 3b, why do the authors select the best 100 values independently for each metric on each target instead of plotting the whole distribution? This selection process should appear in the figure caption or main text, instead of in the appendix.
  2. p. 9, lines 460–462. “Physical plausibility of the interface and local structures” is not directly measured. The evidence consists of prediction confidence and Cα RMSD. These do not directly assess side-chain recovery, clashes, bond geometry, chirality, rotamers, or interface energetics. The reported results support improved predictor confidence and backbone consistency; the broader claim about atomic physical plausibility needs additional measurements.
  3. The mean improvements are concentrated in a few targets in Table 9. For protein Self sequences, CODesign wins against Complexa on 9 targets, loses on 8, and ties on 2. BHRF1 and DerF21 contribute more than the entire net improvement.
  4. Efficiency comparison in Table 2 lack sufficient configuration details, including GPU, precision and compilation. Also, Efficiency comparison in Table 1 is missing.
7 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/5)

We are grateful for your constructive feedback and thoughtful evaluation of our work. Below are our responses to your concerns.

1. Response to the novelty against Proteína-Atomística (W1)

Thank you for raising this important point. We acknowledge that CODesign builds upon the explicit multimodal flow-matching framework and the residue–atom architecture of Proteína-Atomística. However, our work goes beyond directly applying this architecture to a new task. We introduce several task-specific extensions to address the challenges of target-conditioned all-atom binder co-design:

  1. New Task: Target-Conditioned All-Atom Binder Co-Design. Proteína-Atomística primarily focuses on unconditional de novo protein generation, whereas CODesign addresses the more challenging task of generating protein binders conditioned on either protein or small-molecule targets. This requires explicitly modeling binder–target interactions rather than only the internal distribution of a protein. Unlike Complexa, which models target latent representations, we directly let the designed binder and target interact within Euclidean space. Therefore, we incorporate binder–target geometric features into the joint pair representations and pair-biased attention. We further introduce a binder–target distance supervision objective to guide the global positioning of generated binders. These designs enable a unified framework for both protein- and ligand-conditioned binder generation.

  2. New Dataset: CODIMER. While Proteína-Atomística introduces consistency distillation for monomeric protein generation, we extend this principle to protein complexes. Specifically, we construct CODIMER from 191,836 reference dimers collected from AFDB high-confidence heterodimers, HumanPPI, and PINDER. For each complex, we jointly redesign both chains using ProteinMPNN, co-fold the redesigned sequences using ESMFold2-Fast, and select confidence-qualified sequence–structure pairs based on complex-level structural consistency. This yields 104,528 consistency-distilled dimer pairs, providing sequence–structure supervision that explicitly captures intermolecular interfaces and complex-level consistency. We have also proved the effectiveness of CODIMER's complex-level consistency distillation (Reviewer qxm7|Response 1).

  3. New Methodology: Consistency-Aware Joint Resampling (CAR). This is our main algorithmic contribution beyond Proteína-Atomística. Unlike standard stochastic remasking, CAR leverages the model's intrinsic inverse-folding capability to identify residue identities that are incompatible with the evolving predicted backbone. It selectively resamples these residues together with their local atomic states, while incorporating the revised sequence into self-conditioning so that subsequent backbone predictions can also adapt. Importantly, CAR obtains its correction signal directly from the generator, without requiring external folding or inverse-folding models or reward-guided search.

In summary, CODesign extends the explicit all-atom co-design paradigm from unconditional monomer generation to target-conditioned binder design, with consistency addressed at three complementary levels: target-aware joint modeling, complex-level data distillation, and trajectory-level correction. We believe these contributions constitute a meaningful extension beyond Proteína-Atomística.

AuthorsOpenReview ↗

Rebuttal by Authors (2/5)

2. Response to computational efficiency (W2 & Q4)

Thank you for pointing out the computational overhead introduced by explicit atom modeling. We additionally measured the per-sample generation time on protein targets and investigated the performance–efficiency trade-off under reduced sampling steps on both protein and small-molecule targets.

All methods were evaluated on a single NVIDIA GeForce RTX 4090. CODesign and Complexa were run using FP32 with TF32 enabled for matrix multiplications, without FP16/BF16 autocast or torch.compile. For other baselines, we followed the default precision and compilation settings in their respective official repositories.

The per-sample generation times on protein targets are summarized below:

MethodTime (s) ↓
RFDiffusion317.64
PXDesign8.51
Protpardelle-1c5.25
Complexa8.52
CODesign (1,000 global updates)28.41
CODesign (400 global updates)11.89

The default CODesign sampling configuration uses 1,000 global updates, with 500 active updates for the backbone and sequence modalities and 1,000 updates for local atomic refinement. We further investigated reducing the number of global updates while maintaining the same modality schedules. In particular, using 400 global updates (200 backbone updates, 200 sequence updates, and 400 local atom updates), CODesign achieves 9.47 mean unique successes with a generation time of 11.89 seconds per sample.

Notably, this reduced-step configuration retains a higher mean number of unique successes than Complexa (9.47 vs. 8.54) under the Self sequence setting, representing a 10.9% improvement. Meanwhile, its inference time is substantially faster than RFDiffusion3 (11.89s vs. 17.64s). Although CODesign is not the fastest approach, these results demonstrate that reducing the sampling budget substantially improves its efficiency while preserving competitive binder design performance.

We further evaluated the same 400-update configuration on the four small-molecule targets. The per-target and mean unique successes are summarized below:

MethodSAM ↑OQO ↑FAD ↑IAI ↑Mean ↑
Complexa3.203.005.808.405.10
CODesign (1,000 global updates)5.8037.2021.4018.6020.80
CODesign (400 global updates)5.0030.0017.6716.0017.17

On small-molecule targets, the 400-update configuration outperforms Complexa on all four targets, achieving 3.37× the mean unique successes (17.17 vs. 5.10) at only 1.41× the per-sample generation time (14.50s vs. 10.27s). Thus, the improvement in unique successes substantially exceeds the relative increase in generation time. Moreover, compared with our default 1,000-update configuration, reducing the sampling budget to 400 updates lowers generation time from 39.90 to 14.50 seconds (a 63.7% reduction), while retaining the highest mean unique successes among the evaluated baselines.

AuthorsOpenReview ↗

Rebuttal by Authors (3/5)

3. Response to the distribution plotting (Q1)

We apologize for any confusion. Our intention was to investigate how CODIMER improves the quality of generated designs with respect to the metrics relevant to our success criteria. Specifically, for each target and each metric, we plotted the 100 most favorable values independently (i.e., the lower 100 values for ipAE and scRMSD, and the higher 100 values for pLDDT). This visualization focuses on the higher-quality portion of the generated samples, making the differences between the two variants easier to inspect without the lower-quality tail dominating the displayed range. We agree that this selection should have been explicitly stated in the figure caption, and will clarify it in the revised manuscript.

To complement this visualization, we additionally calculated the mean values of all generated samples, without any truncation. The results are as follows:

VariantipAE (Å) ↓pLDDT ↑Binder scRMSD (Å) ↓Complex scRMSD (Å) ↓
CODesign w/o CODIMER19.1550.8110.8917.515
CODesign17.6840.8260.7546.322

The full-sample statistics consistently favor CODesign trained with CODIMER across all four metrics. Specifically, CODIMER reduces the mean ipAE by 1.471 Å, increases pLDDT by 0.015, reduces binder scRMSD by 0.137 Å, and reduces complex scRMSD by 1.193 Å. These results further demonstrate that CODIMER improves both predicted interface confidence and structural consistency across the generated samples, rather than merely improving the selected high-quality portion shown in Figure 3b.

Reviewer_iPSMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CODesign, a conditioned all-atom protein binder design framework. Its main contributions include CODIMER, a consistency-distilled dimer dataset constructed using ProteinMPNN and ESMFold2-Fast; a multimodal joint flow model that explicitly couples discrete residues with continuous coordinates; and the CAR process, which leverages the model’s intrinsic inverse-folding capability to iteratively correct sequence–structure incompatibilities during sampling. On the protein-conditioned protein binder design benchmark, CODesign achieves the highest Unique Success score.

Critical strengths
  1. This paper presents an important solution to the sequence–structure consistency problem. The CAR mechanism addresses potential sequence–structure inconsistencies during the codesign process. CAR leverages inverse-folding capability along the design trajectory while simultaneously revising the sequence and resetting the local all-atom state. This capability enables CODesign to achieve very strong performance on the in silico binder design task.
  2. CODIMER extends the concept of consistency distillation to complexes, further improving the performance of CODesign.
Critical weaknesses
  1. The training data used by CODesign come from multiple sources, including real data from PDB and synthetic data from AFDB. For the benchmarks used in the experiments, potential homology between the training and test data is not explicitly discussed in the manuscript. If the training set contains a large number of structures similar to those in the test set, the benchmark scores could be substantially inflated, making it difficult to assess the true performance of CODesign.
  2. The manuscript lacks comparisons with several representative baselines, such as PXDesign, ODesign, and SeedProteo. In particular, some baselines, such as PXDesign, may achieve higher scores on the conditioned-binder design benchmark than the baselines currently included.
Questions for authors
  1. Could the authors provide specific details or ablation studies regarding the CAR hyperparameters? In particular, the manuscript appears to lack an explanation for the choice of the resampling fraction and the setting of ρ=0.2\rho = 0.2. Since CAR is a newly proposed method in this work, could the authors provide a sensitivity analysis of these parameters to give readers better intuition about their effects?
  2. For the protein targets and ligand targets used in the benchmark, were the training data deduplicated or filtered to remove homologs of the benchmark targets? How does CODesign ensure that there is no leakage from the training set into the test set?
  3. Could the authors analyze how the substitutions proposed by CAR are retained in the final generated sequences, either on a case-by-case basis or across the benchmark? Are these substitutions overwritten during subsequent generation steps, or are they likely to persist until the end and improve the quality of the generated sequences?
5 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/4)

We appreciate the time and effort you have dedicated to reviewing our paper. Below are our responses to your concerns.

1. Response to the homology between the training and test data (W1 & Q2)

Thank you for raising this important concern. We conducted a target-wise overlap analysis of all 658,528 training examples from Teddymer, PDB, and CODIMER. For each benchmark target, we flagged a training example if any of its protein chains has sequence identity > 50% or TM-score > 0.5 to the target, or if their PDB IDs overlap. The results are summarized below:

TargetTeddymerPDBCODIMERTotalOverlap Ratio
IFNAR236857404650.0706%
BHRF101428420.0064%
BBF14511310740.0112%
DerF2101010.0002%
TrkA1,5221704702,1620.3283%
PD14,5333,6809,92018,1332.7536%
Insulin52190710.0108%
DerF72120230.0035%
PDL14,1533,6849,76917,6062.6735%
IL7RA066381040.0158%
CrSAS6541100.0015%
Claudin17223320.0049%
VEGFA52129550.0084%
SpCas900000.0000%
SC2RBD0339123510.0533%
CbAgo00000.0000%
CD451,00710181,0350.1572%
BetV1294101554500.0683%
HER2_AAV1151170.0026%

For 15 of the 19 targets, the overlapping examples constitute less than 0.1% of the training corpus. Notably, CODesign achieves substantial improvements over Complexa on targets with very limited detected overlap. The corresponding Self-sequence results from Table 9 are shown below:

TargetOverlap RatioComplexa: Unique Successes ↑CODesign: Unique Successes ↑
BHRF10.0064%22.0043.60
DerF210.0002%20.6036.40
CrSAS60.0015%6.8014.20
Insulin0.0108%12.2016.40
IL7RA0.0158%4.607.20

These results show that CODesign's substantial performance gains are not restricted to targets with extensive training-data overlap. We also clarify that our Teddymer, PDB, and PLINDER datasets are the same as those used by Complexa, so overlap in these shared training sources is not specific to CODesign.

We are currently retraining on the overlap-filtered dataset. Given the substantial computational cost of full retraining and our current resource constraints, the results are not yet available during rebuttal. We plan to include the completed results in the camera-ready version.

AuthorsOpenReview ↗

Rebuttal by Authors (2/4)

2. Response to missing baselines (W2)

Thank you for suggesting these comparisons. We would like to clarify that PXDesign is already included in Table 1, with detailed results in Appendix F.2. Following your suggestion, we additionally evaluated ODesign under the same protein-target benchmark protocol. The mean unique successes are summarized below:

MethodSelf ↑MPNN@1 ↑MPNN@8 ↑
PXDesign—4.597.01
ODesign—2.496.80
CODesign9.8410.3218.81

CODesign outperforms both PXDesign and ODesign under the matching MPNN@1 and MPNN@8 settings. It also achieves 9.84 mean unique successes using its directly generated sequences, without ProteinMPNN redesign.

For SeedProteo, we could not obtain publicly released inference code and pretrained weights for evaluation under the same benchmark protocol, preventing a direct comparison.

AuthorsOpenReview ↗

Rebuttal by Authors (3/4)

3. Response to ablation studies regarding the CAR hyperparameters (Q1)

Thank you for this insightful suggestion. The resampling fraction is the parameter ρ\rho in Eq. (31). It limits the number of currently unmasked residues that CAR can revisit at each intervention. Thus, ρ\rho controls the trade-off between correcting incompatible choices and preserving the current sequence–structure trajectory. Among eligible positions, CAR selects those with the largest preference gaps. A moderate resampling fraction therefore focuses corrections on the strongest model preferences for replacement while avoiding extensive simultaneous resets of sequence and local atomic states. The default ρ=0.2\rho=0.2 is intended as a moderate correction budget, rather than a claim of global optimality.

To investigate the effects of the CAR hyperparameters, we conducted a fixed-seed sweep over the backbone-time threshold, resampling fraction, and check interval, evaluating all 54 combinations on the 19 protein targets. Each entry below reports the mean unique successes using Self sequences:

Backbone-time thresholdResampling fractionCheck interval = 5Check interval = 10Check interval = 20
0.700.108.539.379.00
0.700.208.689.429.32
0.700.408.639.059.47
0.750.108.638.798.79
0.750.208.638.849.32
0.750.408.848.849.42
0.800.108.429.119.47
0.800.209.009.009.53
0.800.409.008.899.68
0.850.1010.379.588.89
0.850.209.639.589.79
0.850.409.639.959.84
0.900.108.589.479.05
0.900.209.479.429.58
0.900.409.329.429.42
0.950.109.749.959.47
0.950.209.749.1610.11
0.950.409.799.1110.16

The results reveal several trends:

  • Later intervention generally improves performance. Averaging over the other settings, backbone-time thresholds of 0.85 and 0.95 both achieve approximately 9.70 mean unique successes, compared with 8.90–9.12 for thresholds of 0.70–0.80. This supports performing compatibility assessment after the backbone has become more resolved.
  • Longer check intervals are particularly beneficial when intervention begins earlier. For thresholds of 0.70–0.80, increasing the interval from 5 to 20 improves the average score from 8.71 to 9.33. This suggests that allowing the coupled sequence–structure trajectory to evolve between interventions helps avoid repeated corrections based on less-resolved geometry.
  • A moderate resampling fraction provides an effective correction budget. Averaging over the other settings, the scores are 9.18, 9.35, and 9.36 for ρ=0.1\rho=0.1, 0.20.2, and 0.40.4, respectively. Increasing the fraction from 0.1 to 0.2 improves performance, while increasing it further to 0.4 yields a similar average score. These results support ρ=0.2\rho=0.2 as a balanced choice that allows sufficient correction while preserving more of the current sequence–structure state.
Reviewer_qxm7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

CODesign extends Proteína-Atomística to target-conditioned all-atom protein binder co-design, jointly generating sequences, Cα backbones, and local atomic coordinates. It introduces CODIMER, a dataset of 104,528 dimer sequence–structure pairs constructed through joint sequence redesign, co-folding, and filtering, and consistency-aware joint resampling (CAR), which uses the generator’s conditional predictions to revisit potentially incompatible residue identities and local atomic states. Evaluation covers 19 protein targets and four small-molecule targets over five sampling seeds. CODesign achieves the highest mean number of unique in silico successes among the evaluated baselines. Protein-target ablations show an increase from 5.54 to 9.46 with CODIMER, and to 9.84 with CAR. Experimental binding performance remains untested.

Critical strengths
  1. Substantial empirical benefit from CODIMER. With CAR disabled, adding CODIMER increases mean unique successes from 5.54 to 9.46, with improvements on 12 of 19 protein targets. This is the paper’s strongest empirical result and supports the practical value of the proposed data pipeline.
  2. Competitive performance across two target types. The integrated system achieves the highest mean unique successes among the evaluated baselines on both protein- and small-molecule-target benchmarks. CAR also operates without an external folding or inverse-folding model, adding little generation overhead.
Critical weaknesses
  1. The contribution of consistency distillation is not sufficiently isolated. CODIMER extends the redesign–refolding–filtering strategy established by CDDB to dimers, making its clearest contribution a complex-specific data resource. Although the ablation supports the benefit of incorporating CODIMER, it does not distinguish consistency distillation from increased data coverage, source composition, or quality filtering. A source- and size-matched undistilled-data control would materially strengthen attribution of the improvement to the proposed consistency pipeline.
  2. CAR’s specific selection mechanism lacks a controlled comparison. CAR increases mean unique successes from 9.46 ± 0.65 to 9.84 ± 0.25, while generation time rises from 27.96 to 28.41 seconds. However, the ablation only enables or disables CAR on top of a base sampler that already includes stochastic remasking. It does not compare incompatibility-based selection with random position selection under otherwise matched joint-resampling settings and inference budgets. The results therefore support a modest benefit from CAR as a whole, but do not establish the contribution of its selection rule.
Questions for authors
  1. Clarification—training-budget matching. Appendix F.4 states that the CODIMER ablations use the same training hyperparameters. Can you confirm that the variant without CODIMER also receives the additional 10,000 second-stage updates with global batch size 320, and explain how CODIMER’s sampling probability is redistributed?
  2. Additional evidence—dataset attribution. Are results available using a source- and size-matched set of undistilled complexes with comparable quality filtering? This would help distinguish the benefit of redesign and co-folding from data coverage and selection effects.
  3. Additional evidence—CAR attribution and uncertainty. Are paired per-seed results available, together with a control that randomly selects eligible residues while matching the joint resets, self-conditioning procedure, resampling schedule, and inference budget? These would clarify the uncertainty of the gain and the value of incompatibility-based selection.
4 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/3)

We sincerely appreciate your constructive comments on our work. Below are our responses to your concerns.

1. Response to isolating the contribution of consistency distillation (W1 & Q2)

Thank you for this insightful suggestion. To better isolate the contribution of the proposed consistency distillation pipeline, we conducted an additional controlled experiment using source- and size-matched undistilled data.

Specifically, for each of the 104,528 samples in CODIMER, we retrieved its corresponding original input complex before the consistency-distillation pipeline and directly replaced the distilled sample with this original complex. Thus, the undistilled variant uses exactly the same set of reference complexes as CODIMER. This ensures that the two variants have identical dataset sizes and source compositions, differing only in whether the consistency-distillation pipeline is applied. Starting from the same first-stage checkpoint, we then performed the second-stage training using the same training configuration and evaluated both variants on 19 protein targets across five random seeds. All evaluations were conducted using self-generated sequences (Self).

The results are summarized below:

VariantUnique Successes ↑scRMSD (Å) ↓scRMSD ≤ 2 Å (%) ↑
Base (without CODIMER)5.54 ± 0.488.01 ± 0.0734.45 ± 0.62
With CODIMER (undistilled)6.73 ± 0.657.80 ± 0.1236.52 ± 0.90
With CODIMER9.46 ± 0.657.03 ± 0.0441.49 ± 0.54

The results demonstrate that consistency distillation is the primary source of CODIMER's performance gain. While incorporating the matched undistilled complexes improves the mean unique successes from 5.54 to 6.73, applying the consistency-distillation pipeline further increases it to 9.46 (+40.6% over the undistilled variant), accounting for approximately 70% of the total improvement over the base model. The improvement is also reflected in sequence-structure consistency: CODIMER reduces the mean scRMSD from 7.80 Å to 7.03 Å and increases the percentage of designs with scRMSD ≤ 2 Å from 36.52% to 41.49%. Since both variants use exactly the same underlying reference complexes, dataset size, and source composition, these results indicate that the substantial performance gain primarily stems from the consistency-distillation process rather than simply incorporating additional training data.

AuthorsOpenReview ↗

Rebuttal by Authors (2/3)

2. Response to comparing incompatibility-based selection with random position selection (W2 & Q3)

Thank you for pointing out the importance of isolating the contribution of CAR's residue selection mechanism. Following your suggestion, we conducted an additional controlled experiment by replacing CAR's incompatibility-based selection with random selection.

Specifically, at each CAR intervention, instead of selecting eligible residues with the largest preference gaps, we randomly sampled the same number of eligible residues. We kept the joint resets, self-conditioning procedure, resampling schedule, and inference budget unchanged. Thus, the only difference between the two variants is the criterion used to select residues for resampling.

We evaluated both variants using self-generated sequences across five random seeds. The results are shown below:

SeedRandom Joint ResamplingCAR
09.789510.1579
19.42119.7895
29.21059.4737
38.78959.8947
48.78959.8947
Mean9.20009.8421

CAR consistently outperforms random selection across all five seeds, improving the mean unique successes from 9.20 to 9.84 (+7.0%). This controlled comparison thus underscores the value of incompatibility-based selection mechanism introduced in CAR.

AuthorsOpenReview ↗

Rebuttal by Authors (3/3)

3. Response to clarification of the training budget (Q1)

Yes. We confirm that the variant trained without CODIMER also undergoes the complete second-stage training, consisting of 10,000 additional updates with a global batch size of 320, using the same training hyperparameters as the CODIMER variant.

The only difference is the data mixture: the sampling probability originally assigned to CODIMER (0.19) is redistributed to PDB binder–target samples. Specifically, their sampling probability increases from 0.10 to 0.29, while the probabilities assigned to the other data sources remain unchanged.

Thank you for the additional experiments and clarifications.

  • The comparison using the same reference complexes under a matched training budget addresses my main concern about CODIMER: it provides convincing evidence that the distillation pipeline adds value beyond including more dimer data.
  • The CAR comparison also addresses my concern about residue selection. Ranking eligible residues outperforms random selection in all five paired seeds, although the overall gain from enabling CAR remains modest (9.46 to 9.84 mean unique successes).

I value the resulting dataset and the evidence for CAR, but my overall recommendation remains unchanged.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.