acceptodds
Withdrawn from ICLR 2027

Composing Co-Folding Models with Feynman–Kac Correctors

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Feynman–Kac correctors (FKCs) compose pretrained diffusion models by simulating the sum of their scores and correcting the samples with importance weights and sequential Monte Carlo resampling. Diffusion-based all-atom structure predictors are natural candidates for such composition: no single predictor is reliably best, and pooling their outputs helps sampling more than selection because confidence scores are not comparable across models. We introduce FKCFold, which adapts FKCs to independently trained co-folding models that differ in architecture and tokenization but share the same all-atom denoiser parameterization. After each model encodes the complex with its own trunk, a single sampler evolves a population of candidate structures: at every step, all models denoise the same noisy structure, the sampler follows their normalized product of experts, and Feynman–Kac weights penalize disagreement between models. A common all-atom coordinate space, per-step rigid alignment of each prediction, native noise ranges for each model, and a correction adapted to the shared AlphaFold 3-style sampler make this composition well posed. FKCFold requires no training. Across multiple benchmarks, FKCFold combining Protenix-v2, OpenDDE and ESMFold2 consistently matches or outperforms each of the models it combines, with the largest gains on antibody–antigen complexes.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
biomolecular structure prediction, co-folding models, diffusion models, Feynman–Kac correctors, product of experts, sequential Monte Carlo

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_nEdiOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This manuscript proposes an ensembling approach for combining the diffusion trajectories of multiple co-folding models through Feynman–Kac correctors to improve the resulting structure over their individual predictions. It focuses primarily on antibody-antigen modeling, where presumably the greatest gains are to be hard, and shows marginal improvement over a baseline "best of all" method that simply picks the best sample from the ensemble.

Critical strengths
  • The approach is interesting and leverages the fact that the diffusion trajectories of most co-folding models occur in the same space to use FKC to improve the overall accuracy of the final predicted structures. From a conceptual standpoint it's a neat idea and may lay the groundwork for further developments in the future.
  • Experimental design appears sound and includes a broad range of Ab-Ag benchmarks that comprehensively assess performance.
Critical weaknesses
  • The gains are very marginal, particularly with rest to the best-of-all baseline. Given that the baseline also uses self-ranked confidence, it's unclear that the significantly more complex machinery introduced here justifies the cost.
Questions for authors
  • Related to my concern about the paper's primary weakness, what is the computational cost of doing FKC across multiple co-folding trajectories vs. a single co-folding model? Given the inference-time scaling behavior of Ab-Ag prediction for co-folding models, this cost is a critical aspect of the utility of this approach.
  • Why was sampling done without replacement for the "best of all" baseline? If the intent is to emulate the 25 sample budget of individual models, should it not be with replacement?
Reviewer_AKGvOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces FKCFold, a training-free, inference-time framework for ensembling independently trained diffusion-based co-folding models. The method combines the denoiser predictions of Protenix-v2, OpenDDE, and ESMFold2 within a shared sampling process. To handle differences across models, it uses atom mapping, rigid coordinate alignment, and noise-dependent model activation. It further incorporates Feynman–Kac correction (FKC), which reweights and resamples particles based on disagreement among the models' denoising predictions.

The authors evaluate FKCFold on four benchmark panels covering protein–protein and antibody–antigen complexes, including FoldBench, 2026ARK-AB, and PXM-2025. The results show improvements over individual models, particularly on FoldBench antibody–antigen complexes. The paper also includes comparisons with independent prediction pooling, candidate generation and selection analyses, model-removal ablations, and inference-time measurements to assess the effectiveness and practicality of the proposed framework.

Critical strengths
  1. Practical inference-time ensemble framework. The paper presents a practical approach to combining independently trained diffusion-based co-folding models without additional training. In particular, handling different atom representations, coordinate frames, and noise ranges makes the framework applicable to heterogeneous models.

  2. Relatively comprehensive and transparent experimental evaluation. The authors evaluate the method across multiple benchmarks, including comparisons with individual models and independent pooling, model-removal ablations, candidate generation and selection analyses, and inference-time measurements. The improvements on antibody–antigen complexes are particularly interesting. I also appreciate that the authors report inconclusive and negative results, which makes the empirical analysis more convincing.

Critical weaknesses
  1. Unclear contribution of Feynman–Kac correction beyond simple denoiser ensembling. The central sampling update is essentially a weighted average of aligned denoiser predictions from independently trained models. While FKC has been studied in prior work, the paper does not provide a direct comparison against the same shared mean-denoiser ensemble without FKC resampling. This makes it difficult to determine how much of the improvement actually comes from FKC rather than straightforward inference-time ensembling.

  2. Limited evidence for the benefits of the full three-model framework. Although the method improves over individual models, its gains over independent pooling are less conclusive, with confidence intervals including zero across all four benchmarks under the default selector. The model-removal experiments also suggest uneven contributions from the three models, with Protenix-v2 showing the clearest marginal contribution. This raises questions about whether the full three-model ensemble is necessary and where the improvements actually come from.

Questions for authors
  1. How much does FKC actually contribute beyond the shared mean-denoiser ensemble? Could the authors report an ablation without FKC weighting and resampling under identical sampling and selection settings? Could they also provide FKC-related statistics, such as resampling frequency, particle weight distributions, and ESS trajectories?

  2. Why is there a relatively large discrepancy between the original OpenDDE results on FoldBench Ab–Ag (70.0% ranked success and 81.9% oracle success [1]) and those reported in this paper (62.0% and 73.1%)? Are the checkpoint versions, recycling cycles, diffusion steps, MSA/template settings, candidate counts, and evaluation protocols consistent with the original work?

[1] OpenDDE: Folding, Reasoning, and Scaling with Open-source Drug Discovery Engine, Figure 2. arXiv:2607.03787.

  1. Given that the model-removal experiments suggest uneven contributions from the three models, why is the full three-model ensemble preferable to the best two-model combination? Would the three-model ensemble still provide meaningful improvements under matched computational budgets and consistent selection settings?

  2. Why are the improvements particularly pronounced on FoldBench Ab–Ag compared with the other benchmarks? Could the authors provide representative case studies showing where coupled sampling succeeds but individual models or independent pooling fail? What are the training-data cutoffs of the three models, particularly the antibody-specific checkpoints, and how much sequence or interface similarity exists between their training data and the evaluated benchmarks?

Reviewer_7brROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes to combine an ensemble of co-folding models for biomolecular structure prediction instead of using single models. The models are combined during generation, at each denoising step, all models denoise the same candidate structures and their predictions are combined following Feynman-Kac correctors for diffusion models (Skreta et al., 2025). Applying this method requires no training. The authors show that the ensemble of Protenix-v2, OpenDDE and ESMFold2 achieves better accuracy than each individual model. This improvement is shown on protein-protein and antibody-antigen interfaces from FoldBench, 2026ARK-AB and PXM-2025.

Critical strengths
  • The method doesn't require training and can be applied to any diffusion based co-folding model, so it can incorporate new models as they are released.
  • The improvement in accuracy is consistent across benchmarks on a relevant task (protein-protein and antibody-antigen interfaces), and confidence intervals are provided for the estimates.
  • Coupling the models during generation works better than the most basic baseline of pooling the predictions of the individual models.
  • The evaluation is detailed: model removal ablations show the contribution of incorporating each model, and timings are reported.
Critical weaknesses

Baseline ablation

  • The FK correction, which is in the title, is defined as optional (Sec. 3.3), without it, I believe the method reduces to an average of the models' predictions (?). The paper does not clarify whether or how often the correction is used in the reported results. An ablation comparing the method with and without the FK correction would be necessary. Otherwise, the paper's main claim, the use of Feynman-Kac correctors, is weakly supported.

Novelty

  • Ensembles of models tend to improve performance in most tasks, at the expense of extra compute. The novelty is therefore somewhat limited, especially if the benefit of the FK criterion is not shown vs a plain average.
Questions for authors
  • Is the Feynman-Kac correction enabled in the reported results? If so, how often is resampling triggered?

  • Would it be possible to include an ablation that compares FKCFold with and without the (optional) Feynman-Kac correction, i.e. against a plain per-step average of the models' predictions? Since the FK correction is the main methodological contribution and appears in the title, this comparison would show whether it adds to simple averaging.

If these points are addressed, I am happy to increase my score.

Reviewer_sej1OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces FKCFold, a training-free method for composing pretrained co-folding models during inference. Feynman–Kac correction addresses the fact that following an average of the component models’ scores does not generally sample from the intended product-of-experts distribution.

Experiments combine Protenix-v2, OpenDDE, and ESMFold2 on protein–protein and antibody–antigen benchmarks, showing improvements over individual component models at matched candidate counts. Comparisons against independent pooling are less conclusive, and an advantage at matched computational cost is not established.

Critical strengths
  • A practical approach to composing heterogeneous predictors without retraining. Reconciling atom layouts, rigid-coordinate frames, and supported noise ranges is a concrete contribution. The disagreement-based weight can also be computed from existing denoiser outputs without additional denoiser evaluations.

  • Evidence that coupled generation can improve candidate quality. Table 1 holds a single confidence module fixed while comparing standalone and coupled candidate sets. These controls provide meaningful evidence that some improvements arise from generation rather than exclusively from changing the selector. Oracle coverage also improves over individual models at matched candidate counts.

Critical weaknesses
  1. The contribution of the Feynman–Kac correction is not isolated. The most direct baseline is the identical shared-denoiser sampler with the FK importance weighting and resampling disabled, keeping the component models, alignment, activation schedule, particle count, and final selector fixed. I could not find this comparison. Consequently, the reported improvements do not establish whether the correction contributes beyond averaging the models’ denoised predictions along a shared trajectory. This is particularly important because Appendix A acknowledges that alignment, atomwise weight normalization, changing active models, and the discrete sampling dynamics depart from the assumptions of the exact FK construction.

  2. The comparisons do not establish an advantage at matched inference cost. Matching the number of candidates does not match computational cost when each coupled trajectory evaluates multiple denoisers. Table 12 reports 248.9 seconds for five coupled candidates, compared with 271.1 seconds for fifteen independently generated candidates. Nevertheless, the main pooled baseline evaluates 25 candidates subsampled from the 75 available independent predictions. The relevant practical comparisons are larger-(N) sampling from individual components and independent multi-model pooling at the same compute budget. Figure 3 varies candidate counts but does not resolve this issue. The authors acknowledge this limitation, but it leaves the practical benefit over a strong independent-sampling alternative uncertain.

  3. The robustness of the selected configuration remains uncertain. The stored held-out FoldBench analysis in Table 11 does not reproduce the full-panel gains. With the combined selector, FKCFold obtains 66.7% versus 72.2% for Protenix-v2 on protein–protein interfaces, and 55.0% versus 65.0% on antibody–antigen interfaces. These subsets are small, and the authors appropriately describe the analysis as a sensitivity check. Nevertheless, the reversal makes it important to clarify how the correction and selection settings were chosen and how well they transfer to targets not used during development.

Questions for authors
  1. What happens when FK correction is disabled? Please compare the same shared-denoiser sampler with and without importance weighting and resampling, holding all other settings and the final selector fixed. How frequently does resampling actually occur?

  2. Does the method outperform independent sampling at matched compute? Please compare against larger-(N) sampling from the individual components and independent multi-model pooling, reporting accuracy versus measured GPU time and component-wise denoiser evaluations. Include the cost of confidence evaluation and final selection.

  3. How much of the improvement survives a common final selector? What are the results when standalone and pooled candidates receive the same three-model confidence evaluation and density-based ranking as FKCFold? Appendix I also reports that the same coupled structures achieve 75.0% success with stored confidence values but 50.0% after common rescoring; what causes this discrepancy, and how sensitive are the main conclusions to the scoring protocol?

  4. How were the correction and selection settings chosen?

  5. Could the relationship to other inference-time composition and steering methods be clarified? FKSFold is relevant prior work on confidence-guided FK steering, although its molecular-glue setting differs from this paper’s experiments. SoupFold is a closely related contemporary approach to representation-level composition. A discussion would help position the contribution.

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

Related papers

Discussion (0)

Sign in to comment.