acceptodds
Under review as a conference paper at ICLR 2027

Co-folding with a Soup of Representations

53 views

Abstract

Co-folding models such as AlphaFold3, Protenix, ESMFold2, and OpenDDE have advanced rapidly, yet no single model consistently performs best across all biomolecular complexes. In this paper, we show that their pair representations encode complementary information that can be transferred across models to improve structure prediction. We introduce SoupFold, which combines pair representations from multiple co-folding models in a common representation space and generates structures from the combined representation. Importantly, SoupFold does not retrain the co-folding models and learns only simple mappings to transfer representations across models. We evaluate SoupFold on antibody-antigen, protein-protein, protein-ligand, molecular glue, GPCR, and oligomeric complex prediction using AlphaFold3, Protenix, ESMFold2, and OpenDDE. By combining representations across models, SoupFold improves over individual co-folding models across the considered benchmarks.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein, protein structure prediction, protein co-folding prediction

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 4.00 over 3 reviews, above 99% of the others.
Reviewer_WRC7OpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces SoupFold, a representation-level ensembling framework designed to improve biomolecular complex structure prediction by integrating complementary information from multiple pretrained co-folding models. The authors observe that the pair representations generated by independently trained structure prediction models exhibit substantial alignment while retaining distinct structural information, particularly when the models produce different predictions. Motivated by this observation, SoupFold uses lightweight multilayer perceptrons to map pair representations from several peer models into the representation space of a selected anchor model. The mapped representations are then combined through weighted averaging, and the resulting representation is passed to the anchor model's diffusion module for structure generation. Importantly, the framework does not require retraining the underlying co-folding models. The authors evaluate the method using AlphaFold3, Protenix, ESMFold2, and OpenDDE across antibody–antigen, protein–protein, protein–ligand, molecular-glue, GPCR, and oligomeric complex prediction benchmarks. SoupFold demonstrates substantial improvements over several standalone models, including an increase in antibody–antigen acceptable DockQ success from 47.65% to 61.76% for Protenix and from 59.41% to 68.24% for OpenDDE. Additional experiments indicate that representation-level ensembling can recover structural interfaces missed by individual models and can outperform conventional sample-level ensembling under the evaluated settings.

Critical strengths

The primary strength of SoupFold is its simple yet effective approach to exploiting complementary information across independently trained co-folding models. Rather than combining final predicted structures or retraining large foundation models, the proposed framework operates directly on intermediate pair representations, providing a flexible strategy for improving existing structure predictors. The use of lightweight representation transfer networks is attractive because it decouples the ensembling mechanism from the expensive process of training new co-folding models. The empirical evaluation is extensive, spanning multiple anchor models, conventional structure prediction tasks, and more challenging higher-order complexes.

Critical weaknesses

The most significant limitation concerns the computational cost and the mechanism underlying the reported performance improvements. Although SoupFold trains only lightweight transfer networks, inference requires executing the trunks of multiple large co-folding models, which may substantially increase runtime, GPU memory consumption, and deployment complexity relative to a standalone predictor. The paper includes a sampling-budget comparison, but it does not fully establish the computational efficiency of the method under matched wall-clock time or total GPU cost across all tasks. Another concern is that the transfer networks are trained to minimize mean-squared error against the anchor representation, which encourages reconstruction of the anchor's features rather than explicitly preserving complementary information. Therefore, the experiments do not fully distinguish genuine complementary-feature integration from representation averaging, denoising, or implicit regularization.

Questions for authors
  1. Mechanism of representation complementarity: Since the transfer networks are optimized to reconstruct anchor-model representations, how can the authors demonstrate that SoupFold preserves genuinely complementary structural information rather than primarily performing representation denoising or averaging? Could controlled experiments with independent checkpoints of the same model, shuffled peer representations, or self-distillation help isolate this mechanism?

  2. Computational cost and efficiency: What are the actual wall-clock inference times, peak GPU memory requirements, and total computational costs of SoupFold compared with individual co-folding models? How do the results change when competing methods are allocated identical computational budgets?

  3. Generalization of transfer networks: The representation mappings are trained on protein-only complexes. How well do these mappings generalize to ligand-containing complexes, modified residues, and other heterogeneous molecular systems? Could the authors analyze transfer accuracy separately for protein–protein, protein–ligand, and ligand–ligand representation entries?

  4. Negative transfer and model selection: SoupFold occasionally degrades performance, such as on the OpenDDE molecular-glue benchmark. Can the authors characterize when peer representations introduce conflicting or misleading signals? Could an adaptive confidence-aware weighting or gating mechanism prevent such degradation?

  5. Comparison with stronger ensembling baselines: The sample-level ensembling baseline selects predictions using confidence scores from different models, which may not be calibrated consistently. Could the authors compare against confidence-calibrated pooling, learned ranking, or compute-matched oracle analyses to better establish the benefit of representation-level integration?

  6. Evaluation selection bias and statistical significance: FoldBench results are reported only for complexes where all four models generate predictions. How many benchmark examples are excluded, and does SoupFold remain effective under a broader evaluation protocol? Could the authors report paired bootstrap confidence intervals or statistical significance tests for the reported improvements?

  7. Scalability and model dependence: How does performance vary with the number and identity of peer models? Is the improvement primarily driven by a particularly strong peer, or does combining multiple weaker peers offer consistent additional benefits? Could the framework be extended to use AlphaFold3 as an anchor model and to evaluate nucleic acid-containing complexes?

5 comments

Dear Reviewer WRC7, we express our deep appreciation for your time and insightful comments. In our updated manuscript, we highlight the changes in blue\color{blue}{\text{blue}}.

In what follows, we address your comments one by one.


W1/Q2. While Figure 10 addresses sampling budgets, what are the wall clock time, peak memory, and total cost, and how do the results change under identical compute budgets?

To address your question, Table R6 reports per-system inference time and peak memory. The additional wall-clock time mainly stems from peer trunks while using the diffusion module only for the anchor model. SoupFold runs model trunks sequentially, bounding peak memory by the largest trunk.

Table R6. Trunk time (s) / diffusion time (s) / peak memory (GiB) for five samples on one AMD Instinct MI300X GPU.

Model500 tokens1000 tokens1500 tokens2000 tokens
AlphaFold316.7 / 7.0 / 3.228.2 / 14.1 / 8.781.0 / 24.0 / 18.6200.3 / 42.8 / 26.5
Protenix11.1 / 9.3 / 5.342.7 / 18.7 / 12.8111.2 / 36.5 / 27.6237.1 / 66.3 / 41.9
ESMFold28.2 / 5.9 / 20.535.0 / 20.2 / 41.498.6 / 43.7 / 78.2213.7 / 85.4 / 137.0
OpenDDE12.2 / 11.5 / 11.255.5 / 28.5 / 36.6156.4 / 59.5 / 77.2355.9 / 91.9 / 140.7
SoupFold (OpenDDE anchor)48.7 / 11.4 / 17.5164.0 / 28.6 / 36.6454.0 / 61.1 / 77.21027.4 / 90.8 / 140.7

We advance the comparison in Figure 10 of the manuscript to consider GPU-time budgets (Table R7). SoupFold retains its advantage over standalone models under comparable compute budgets.

Table R7. FoldBench Ab-Ag accept rate / GPU-h / number of seeds.

Model / SoupFold cost (GPU-h)≈5≈10≈15≈20≈25
AlphaFold347.18 / 4.96 / 449.44 / 9.92 / 850.68 / 14.88 / 1251.36 / 19.84 / 1651.71 / 23.56 / 19
Protenix47.04 / 5.76 / 447.74 / 10.08 / 748.29 / 14.40 / 1048.71 / 20.16 / 1448.84 / 24.48 / 17
ESMFold252.59 / 4.96 / 455.02 / 9.92 / 856.67 / 14.88 / 1258.05 / 19.84 / 1658.94 / 23.56 / 19
OpenDDE58.87 / 5.79 / 359.90 / 9.65 / 560.92 / 15.44 / 861.44 / 19.30 / 1062.11 / 25.09 / 13
Pooled standalone (four models)58.94 / 5.85 / 160.18 / 11.70 / 260.65 / 17.55 / 360.59 / 23.40 / 460.59 / 29.25 / 5
SoupFold (OpenDDE anchor)63.41 / 4.71 / 165.06 / 9.42 / 266.29 / 14.13 / 367.41 / 18.84 / 468.24 / 23.55 / 5

We have modified Section 3.3 and Figure 10 of the updated manuscript to consider GPU-time budgets. We have also incorporated these into Appendix D.7 (Table 15).

wall-clock time or total GPU cost across all tasks

We revisit Table 1 of the manuscript using a single SoupFold seed versus standalone models with comparable budgets (Table R8), including P-P and P-L tasks, which are less sensitive to sampling budgets. SoupFold improves over each anchor with comparable GPU time.

Table R8. FoldBench accept rate / GPU-h / number of seeds.

ModelAb-AgP-PP-L
AF347.2 / 4.96 / 474.4 / 8.00 / 461.4 / 15.52 / 4
Protenix47.0 / 5.76 / 474.4 / 9.64 / 462.5 / 18.16 / 4
+ SoupFold58.7 / 4.52 / 177.4 / 7.65 / 165.5 / 14.19 / 1
ESMFold252.6 / 4.96 / 475.9 / 8.48 / 464.3 / 15.64 / 4
+ SoupFold57.8 / 4.49 / 178.1 / 7.62 / 166.5 / 14.05 / 1
OpenDDE58.9 / 5.79 / 373.8 / 9.90 / 355.1 / 18.39 / 3
+ SoupFold63.4 / 4.71 / 178.8 / 7.95 / 162.6 / 14.76 / 1

W2/Q1. Since transfer networks reconstruct the anchor, how does SoupFold preserve complementary information? Can authors consider controls with independent checkpoints of the same model, shuffled peer representations, or self distillation isolate this mechanism?

(a) Representation complementarity.

Complementary information can survive transfer because the MLP does not reconstruct the anchor exactly. We train a shared small MLP to minimize average MSE across complexes separated from the benchmarks to learn an overall peer-to-anchor alignment. Figure 3(b,left) of the manuscript shows that representations remain less aligned after transfer when the models predict different structures.

Figure 9 of the manuscript provides further evidence that this retained information is useful. Standalone models recover different subsets of native contacts for the same target, while SoupFold combines their representations to recover their union and induces high-quality predictions.

(b) Controlled experiments.

To address your comments, we compare SoupFold with self-distillation and shuffled peer representations. For self-distillation, we average representations from different random seeds of the same model. For shuffled peers, we permute peer token order before transfer. We cannot test independently trained checkpoints of the same model because they are unavailable. Table R9 shows that neither control reproduces SoupFold's gains.

Table R9. FoldBench Ab-Ag accept rates for the control experiments.

AnchorStandaloneSelf-distillationShuffleSoupFold
Protenix47.6548.242.3561.76
ESMFold252.3551.761.1862.94
OpenDDE59.4159.411.7668.24

These controls suggest that SoupFold's gains are not simply artifacts of averaging or perturbing. We have incorporated these into Section 3.4 and Appendix D.11 (Table 19) of the updated manuscript.


Q3. The transfer networks are trained on protein only complexes. How do they generalize to ligands, modified residues, or other heterogeneous molecular systems?

We would like to clarify that our transfer networks are trained on both protein complexes and protein-ligand complexes, but not on nucleic acids. We have revised the “protein-only” description in Appendix A of the updated manuscript to clarify this. The protein-ligand benchmarks in the paper, FoldBench PL, Runs N' Poses, molecular glue, and GPCR, contain ligands and modified residues.

(Q7) Could the framework be extended to evaluate nucleic acid-containing complexes?

To address your question about heterogeneous molecular systems, we evaluate SoupFold on the FoldBench protein-DNA with OpenDDE as the anchor. We present the results in Table R10. SoupFold improves the performance, showing that our transfer networks can benefit protein-DNA prediction.

Table R10. FoldBench protein-DNA accept rates using the highest-confidence prediction among five samples from a single seed.

ModelDockQ ≥ 0.23DockQ ≥ 0.49DockQ ≥ 0.80
AlphaFold375.9563.6145.25
Protenix73.4263.9248.10
ESMFold272.7862.3443.35
OpenDDE77.2262.9739.87
SoupFold (anchor: OpenDDE)79.4366.1442.40

We have incorporated this into Section 3.4 and Appendix D.10 (Table 18) of the updated manuscript, including discussion of training on nucleic-acid complexes as a future direction.


Q4. Can the authors characterize when peer representations introduce conflicting or misleading signals? Could adaptive weighting or gating prevent such degradation?

SoupFold does not guarantee success when the anchor succeeds but all peers fail. Although unsuccessful predictions tend to have smaller representation magnitudes and less influence when ensembled (Figure 3(a, right) in Section 2.2 of the manuscript), this does not fully prevent degradation, as shown in Figure 5(b) of the manuscript.

We explore peer selection and magnitude weighting in Table 4 of the manuscript. With ESMFold2 as the anchor, selecting one peer and weighting by magnitude improves antibody-antigen success at DockQ ≥0.23\geq 0.23 from 62.9462.94\\% to 66.4766.47\\%, relative to averaging all three peers equally. However, the same strategy reduces protein-protein success at DockQ ≥0.80\geq 0.80 from 48.5448.54\\% to 43.8043.80\\%. As discussed in Section 5 of the manuscript, more principled peer weighting can be a future direction for further improvements.

Q5. Can authors consider calibrated pooling, learned ranking, or compute-matched oracle-based comparisonn?

(a) Calibrated confidence and learned ranking.

To address your comment, we first compare SoupFold with confidence-calibrated pooling using isotonic regression [1] and learned ranking using VoroIF-GNN [2]. For calibration [1], we use the held-out dataset SAbDab-22H2, which consists of 99 antibody–antigen complexes and is disjoint from FoldBench.

We present the results in Table R11. SoupFold outperforms these baselines despite their larger sample pool.

Table R11. FoldBench Ab-Ag accept rates for pooled selection and SoupFold.

Selection methodSamplesRanking selection
Raw confidence10060.59
Isotonic calibration10062.6
Learned ranking (VoroIF-GNN)10062.1
SoupFold (OpenDDE anchor)2568.24

We have incorporated this comparison into Section 3.4 (Table 7) of the updated manuscript.

(b) Compute-matched oracle-based comparison.

Table R12 compares pools with matched budgets under oracle selection (highest ground-truth DockQ).

Table R12. FoldBench Ab-Ag oracle accept rates and Foldseek-multimer cluster counts (TM-score threshold 0.800.80).

Prediction poolTrunk passesDiffusion samplesClustersOracle selection
Standalone (AF3, ESMFold2, Protenix-v1, OpenDDE)201001085.88
SoupFold (AF3, ESMFold2, Protenix-v1, OpenDDE as anchors)20100685.88
Standalone + SoupFold202001389.41

Oracle accept rates are identical. Pooled standalone produces more diverse samples from distinct representations, potentially including correct but low-confidence predictions. SoupFold samples from ensembled representations, reducing diversity but more reliably inducing correct predictions (Table R11).

It is noteworthy that combining both pools (200200 samples) raises oracle accept to 89.4189.41\\%. This suggests that varying peer combinations, including standalone as the no-peer case, may broaden coverage. We have incorporated this analysis into Appendix D.9 (Table 17) of the updated manuscript.


Q6-1. How many systems were excluded, and does the conclusion hold under a broader evaluation protocol?

The original FoldBench evaluation excluded two Ab-Ag, five P-P, and 2828 P-L targets. To address your concern about broader evaluation, we rerun the excluded P-P and P-L targets after resolving technical failures. We modify a kernel for out-of-memory issues and modify codes to handle ligand parsing errors.

Two Ab-Ag, two P-P, and four P-L targets remain unscorable because the scorer fails on all predictions. Table R13 reports the additional targets for which both anchor-alone and SoupFold predictions can be scored. SoupFold does not degrade performance on these additional targets.

Table R13. Accept counts on additional FoldBench targets, shown as Standalone / SoupFold.

ModelP-P (n = 3)P-L (n = 24)
ESMFold22 / 213 / 13
Protenix2 / 214 / 15
OpenDDE2 / 214 / 14

Q6-2. Could the authors report paired bootstrap confidence intervals or statistical significance tests for the reported improvements?

To address your comments, we report paired bootstrap 9595\\% confidence intervals for the improvement over each anchor in Table R14.

Table R14. FoldBench accept-rate gains (percentage points), shown as mean gain [paired bootstrap 9595\\% CI].

AnchorAb-AgP-PP-L
Protenix+12.35 [+6.47, +18.82]+4.01 [+1.09, +6.93]+3.23 [+0.19, +6.28]
ESMFold2+7.06 [0.00, +14.12]+2.19 [−0.73, +5.11]+3.42 [+0.57, +6.27]
OpenDDE+8.82 [+4.71, +13.53]+3.28 [+1.09, +5.84]+8.37 [+5.13, +11.79]

We have incorporated statistical analysis into Appendix D.1 (Tables 9) of the updated manuscript.

Reviewer_uSEkOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This manuscript proposes an ensembling approach for co-folding models whereby an anchor model like Protenix is combined with adapted representations from other models before being fed into the diffusion module to produce the final 3D structure. Results are impressive, with consistent gains across modalities, although no single model emerges as the pareto choice across all modalities.

Critical strengths

This is a great paper across all fronts. Well written, clever idea, great implementation, and great results.

Critical weaknesses

No critical weaknesses.

Questions for authors

Why not include Protenix v2 in the list of models, in lieu of PX1, as PX2 appears to be the strongest current co-folding model across the board?

2 comments

Dear Reviewer uSEk, we express our deep appreciation for your time and insightful comments. In our updated manuscript, we highlight the changes in blue\color{blue}{\text{blue}}.

In what follows, we address your comments one by one.


Q. Why the authors use Protenix-v1 rather than Protenix-v2?

We use Protenix-v1 because the v2 weights are no longer publicly available from the official source, and the repository labels them proprietary and confidential. To address your question, we evaluate Protenix-v2 as an anchor exclusively for this rebuttal, using weights from a third-party mirror on Hugging Face. We present the results in Table R4. SoupFold consistently improves performance with the cutting-edge Protenix-v2 as the anchor.

Table R4. FoldBench Ab-Ag results with Protenix-v2 as the anchor.

ModelAcceptMediumHigh
Protenix-v257.6550.5924.12
+ SoupFold64.1253.5326.47

strongest current co-folding model across the board

In addition, we further compare SoupFold with the strongest current co-folding models. We use CoFold Arena (which excludes AF3 for accessibility reasons), where official inputs, samples, and scores are available. Following its protocol, we evaluate 422422 antibody-antigen targets. We use OpenDDE as the anchor and Protenix-v1 and ESMFold2 as peers. SoupFold outperforms all compared models, including latest Mol-1 (Table R5).

Table R5. CoFold Arena Ab-Ag results. Baselines use official leaderboard scores dated October 2929, 20262026. Bold indicates the best result.

ModelRelease dateMean DockQAccept
SoupFold-0.39853.8
Mol-12026-10-270.35948.8
OpenDDE2026-07-060.35546.0
OpenDDE-AbAg2026-07-060.35246.4
Protenix-v22026-04-080.32744.3
ESMFold22026-05-270.26134.1
Protenix-v1-202506302026-02-050.25335.1
IntelliFold-22026-02-070.24934.1
Protenix-v12026-02-050.22431.5
RosettaFold3-01/242025-08-150.19927.3
Chai-12024-09-090.19026.5
Promera2026-06-100.18927.5
Boltz-22025-06-060.17523.9

We have modified Figure 1 of the updated manuscript using Table R5.

Thanks for the additional results--nice work! My score is already 4 so can't raise any further!

Reviewer_ip57OpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The authors present a method for combining multiple competing state-of-the-art protein-ligand cofolding models into an approach that outperforms the base models across a range of benchmarks. The authors begin from the observations that most protein-ligand models share a similar underlying architecture (iterative refinement of pair representation followed by a diffusion module for final structure generation) and that representations are generally less aligned between models when structure predictions may be complementary. The authors build these observations into SoupFold, an approach for combining pair representations across models before feeding into the frozen diffusion module of an "anchor" model. SoupFold-ensembled models uniformly outperform the anchor model across all tested anchor models.

Critical strengths
  1. Well-motivated and timely: co-folding is an important problem in the field and many different approaches have been developed. Exploring whether an informed ensembling method can exploit complementarity between the models makes sense
  2. Comparisons to sample-level ensembling across models and compute-matched scaling for the top-performing single model on antibody-antigen prediction suggests that their pair representation-level ensembling is benefiting from complementary latent information in each model rather than just additional compute or distinct sample-level predictions
  3. Empirical results across multiple benchmarks show consistent improvement of SoupFold-augmented models over the corresponding anchor model
Critical weaknesses
  1. It seems like a "perfect" transfer network that minimizes the MSE loss would result in no additional information over the anchor model, as the transferred pair representations would exactly match those of the anchor model. Do the authors have any intuition as to how important regularization or other methods to mitigate perfect alignment are to the success of the approach? Some discussion of this could help readers understand the discrepancy between the high-level goal of the method (leverage representation-level complementarity across models) and the low-level goal of the loss function (perfect alignment of source and anchor representations)
  2. While informative, some aspects of Figure 3 are unclear. What protein-ligand pairs are being shown? Is each point a single protein-ligand pair or a protein-ligand-prediction model triplet? Caption says similar and dissimilar pairs are defined by DockQ ≥\geq 0.8 and DockQ ≤\leq 0.23, but then what defines "model wrong" and "model right"?
Questions for authors
  • When combining many models with different training protocols, a main concern is data leakage. Can the authors speak to the steps taken or research performed to confirm that none of SoupFold's constituent models were trained on any of the benchmarking data?
  • Mapping each individual pair i,ji,j representation separately seems like it might lose information shared across pairs. Have you all tried any other architectures? e.g. full transformer or lighter-weight alternatives like 2D CNN, transformer with axial attention, etc
2 comments

Dear Reviewer ip57, we express our deep appreciation for your time and insightful comments. In our updated manuscript, we highlight the changes in blue\color{blue}{\text{blue}}.

In what follows, we address your comments one by one.


W1. Perfect reconstruction would make the transferred representation identical to the anchor. Is regularization or other methods needed to preserve complementary information from the peers?

We first clarify that our transfer networks do not perfectly reconstruct the anchor. We train simple MLPs on representations from the co-folding models' training data, separated from the benchmarks (see Table R2 of Q1). The cosine similarity between the transferred and anchor representations is around 0.60.6 (Figure 3(b) of the manuscript).

However, as you mentioned, if reconstruction were perfect, complementary information from the peers could be lost. We conduct a toy experiment assuming extreme over-fitting: a separate large target-wise transfer network is over-fitted for each FoldBench complex exclusively to perfectly reconstruct its anchor representation. Table R1 shows that this reduces SoupFold's gains.

Table R1. FoldBench Ab-Ag results. SoupFold (over-fitted) uses a different transfer network for each complex, over-fitted exclusively to that complex.

ModelAcceptMediumHigh
Protenix47.6538.2422.94
+ SoupFold (ours)61.7654.1223.53
+ SoupFold (over-fitted)50.5940.0022.35
ESMFold252.3543.5324.12
+ SoupFold (ours)62.9451.1824.12
+ SoupFold (over-fitted)51.1841.1822.94
OpenDDE59.4152.3529.41
+ SoupFold (ours)68.2457.0631.76
+ SoupFold (over-fitted)62.9451.7627.65

These results suggest that avoiding over-fitting to the anchor representations helps preserve complementary information from the peers. In our settings, training a small transfer network on data held out from the benchmarks is sufficient to prevent this. We have added this discussion to Section 3.4, line 465, and Appendix D.8 (Table 16) of the updated manuscript.


W2. Which targets are shown in Figure 3 of the manuscript, what does each point represent, and how do "similar/dissimilar" differ from "model right/wrong"?

Figure 3 of the manuscript uses antibody-antigen targets from FoldBench. The two panels in Figure 3(a) make different comparisons:

  • (a, left): Each point shows the CKA between two models’ representations for the same target. Their predicted structures are “similar” if their pairwise DockQ is ≥0.8\geq 0.8 and “dissimilar” if it is ≤0.23\leq 0.23.
  • (a, right): Each point shows one model’s interface representation magnitude for one target. We label predictions as “model right” if their DockQ ≥0.23\geq 0.23 against the ground truth, and “model wrong” otherwise.

Figure 3(b) uses the same definitions after transfer to the ESMFold2 representation space. We have revised the caption of Figure 3 of the updated manuscript.


Q1. Were any of the models trained on the benchmark data?

None of the benchmark structures are used to train the models, as the training and benchmark data are temporally separated. We provide the RCSB PDB release-date cutoffs for training structures and the release dates of benchmark structures in Table R2.

Table R2. RCSB PDB release dates of training and benchmark structures.

Model or datasetPDB release-date cutoff / range
AlphaFold3, Protenix-v1, ESMFold2, OpenDDEBefore September 30, 2021
Transfer networks (ours)Before September 30, 2021
FoldBenchJanuary 13, 2023 to November 01, 2024
Runs N’ PosesSeptember 30, 2021 to January 9, 2025
Molecular glue (MGBench)After September 30, 2021
GPCRDecember 20, 2023 to January 22, 2025

Q2. Were other transfer network architectures tried?

Thank you for suggesting this! To address your comment, we additionally implement a transfer network with a one-layer Pairformer. We present the results in Table R3.

Table R3. FoldBench Ab-Ag results with MLP and one-layer Pairformer transfer networks.

AnchorTransferAcceptMediumHigh
Protenix-v1MLP61.7654.1223.53
-Pairformer64.7152.9422.94
ESMFold2MLP62.9451.1824.12
-Pairformer63.5352.9424.12
OpenDDEMLP68.2457.0631.76
-Pairformer68.2458.2431.76

These results suggest that better transfer-network design may provide additional gains. We have incorporated this into Section 3.4 (Table 6) of the updated manuscript.

Thank you for the detailed reply!

I appreciate the clarifications regarding strategies used to mitigate overfitting (it's clear the original transfer networks didn't overfit, hence the performance gains, so was mainly curious to hear the authors' thoughts on the topic) and the dates associated with temporal cutoffs of the various datasets.

Glad to hear that a small tweak to the transfer network garnered some further improvements. Great work!

Comments on the paper

2 comments
~LunFang2OpenReview ↗

Question about Multi-Representation Fusion Methods

Dear authors,

Thank you for sharing this interesting work. I have a question about the choice of fusion strategy in SoupFold.

After mapping the peer pair representations into the anchor model’s representation space, SoupFold combines them by averaging. Have you also experimented with other fusion strategies, such as concatenation followed by a learned projection, cross-attention between the anchor and peer representations, or other learned fusion methods?

If so, could you share how they compared with averaging in terms of prediction performance and computational cost? If these methods were not evaluated, I would be interested in your thoughts on their potential benefits or practical challenges.

Thank you!

Thank you for your comments and for finding our work interesting!

While we have not experimented with learnable fusion strategies, such methods might capture fusion patterns that further benefit structure prediction. Our results in Figure 6 and Table 4 show that (1) the alignment between anchor and peer representations is associated with performance gains and that (2) simple gating based on this can provide some improvements. Learnable fusion might benefit from capturing such patterns. However, complex learnable fusion may risk pushing representations outside the anchor's representation space, requiring more careful training data curation than our simple transfer networks.

Thank you again for your valuable comment!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.