acceptodds
Under review as a conference paper at ICLR 2027

Guiding Where the Antibody Binds: Amortised Cross-Subsample Particle Steering for Antibody–Antigen Co-Folding

Abstract

Co-folding models have advanced antibody-antigen structure prediction, but a dominant failure mode persists. They often dock the antibody onto the wrong epitope. Docking has two problems. A sampler must produce an acceptable complex at the correct interface, and a verifier must then rank it above the alternatives. We address the first, site coverage, since a verifier cannot select an epitope the sampler never proposed. Drawing more seeds or samples adds variety around the pose the model already favours, but does not relocate it. We introduce CAST (Cross-subsample Amortised Site Tilting), which steers diffusion trajectories toward a binding site using a reward built from a candidate epitope map, training-free and with no reference structure. The sampler approximates an ideal target whose tilt provably changes the site marginal but not the pose at a site. Amortising alignment-subsample latents into one steered pool cuts the cost of exploring them from to for a pool of particles. At a matched structures and conditionings, CAST holds an acceptable pose on of the difficulty-balanced CFA50 targets against for Best-of-, and raises mean max DockQ from to , a gain of . On targets that a -seed run fails to reach CAPRI-medium quality, CAST reaches mean max DockQ at structures and brings of the to an acceptable pose against at the same budget. Best-of- flattens, spending the compute and trailing by . Raising the latent count alongside the samples carries this to , beyond every Best-of- budget. On CFA50 CAST also matches the best Best-of- allocation of structures, at less compute. Such steering is a reallocation, changing little where the base model already succeeds and much where it does not.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
antibody-antigen structure prediction, inference-time scaling, guided diffusion sampling, twisted sequential Monte Carlo, co-folding, reward-guided generation, protein structure prediction

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_isBcOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper targets a key failure in antibody–antigen co-folding: models often repeatedly dock antibodies to the wrong epitope, and additional sampling does not reliably fix this. It introduces CAST, a training-free diffusion-steering method that explicitly explores candidate binding sites while amortizing multiple site hypotheses and MSA conditionings in a single particle pool, reducing exploration cost from O(KN)O(KN) to O(N+K)O(N+K). The authors also provide a reference-free epitope map and a theoretical result showing that site steering preserves the model’s pose distribution conditional on a site.

Empirically, CAST improves mean max DockQ from 0.322 to 0.373 and acceptable-pose coverage from 26/50 to 36/50 on CFA50 at matched sampling budgets, with gains concentrated on difficult targets. On 21 especially hard targets, it succeeds on 18/21 vs. 9/21 for Best-of-N and scales better with additional compute. Comparisons to four steering baselines, ablations, and transfer to a second backbone further support that directed site exploration, rather than diversity alone, drives the gains.

Critical strengths
  1. Novel and well-motivated approach to a consequential failure mode. CAST directly targets site coverage—cases where co-folding models fail to sample the correct epitope at all—rather than relying on additional sampling or pose diversity. The distinction is important because no downstream ranker can recover a correct complex that was never sampled. The training-free formulation and O(N+K)O(N+K) amortization also make the contribution practically meaningful.

  2. Strong, targeted empirical validation. At matched budgets, CAST substantially improves difficult cases (36/50 vs. 26/50 acceptable CFA50 targets; 18/21 vs. 9/21 on the hard set), with ablations and a second-backbone experiment supporting the proposed mechanism rather than generic diversification. This makes the central claim convincing and is the strongest evidence supporting acceptance.

Critical weaknesses
  1. The strongest gains are concentrated on a deliberately difficult regime. CAST provides little improvement on targets that Best-of-N already solves reliably, while the separate 21-target benchmark is explicitly selected for cases where the base model performs poorly. This makes the method compelling as a targeted remedy, but leaves its broader benefit across representative antibody–antigen prediction workloads less established.
  2. Performance remains strongly dependent on candidate-site quality. The oracle-map ablation improves mean max DockQ from 0.373 to 0.484, indicating substantial headroom attributable to site proposal rather than steering itself. Because the practical reference-free map can miss or poorly localize the true epitope, this dependency is a consequential limitation of the end-to-end method.
Questions for authors

Clarification requests:

  1. The overall CAST–Feynman–Kac comparison is not statistically significant (p=0.104p=0.104), although CAST is significantly better on the subset of targets not reliably solved by BoN. Could the authors clarify whether the claimed advantage over alternative steering methods is intended specifically for this difficult regime, rather than overall?

  2. The oracle-map ablation improves mean max DockQ substantially (0.373 → 0.484). Could the authors further quantify how CAST performance varies with candidate-map accuracy, particularly when the true epitope is poorly covered or absent from the candidate set? This would clarify whether candidate-site generation or steering is the primary practical bottleneck.

Suggestion for future work: Improving the reference-free candidate-site map, potentially using stronger epitope predictors, appears to be a promising direction given the large oracle-map headroom.

4 comments

We thank the reviewer for a reading that matches the paper's claims closely. Both weaknesses are properties we agree the paper has.

W1 and Q1 together, where the gain sits and why the pooled Feynman Kac test is not significant.

These are one phenomenon. The effect is concentrated by difficulty group, which Proposition 1 predicts, since the tilt cannot move a pose already at the right site. CFA50 is difficulty balanced, 16 / 13 / 21, and the solved group is reported precisely to show CAST does not damage it. The paired tests are Table 1 and Table 2, with intervals added here (stated plainly in section 4.2).

groupnCAST minus BoNp95% CICAST minus FKp95% CI
not reliably solved29+0.0980.003[+0.043,+0.155][+0.043, +0.155]+0.0820.003[+0.028,+0.146][+0.028, +0.146]
base model solves21-0.0150.99[−0.073,+0.037][-0.073, +0.037]-0.0290.263[−0.070,+0.010][-0.070, +0.010]
all50+0.0510.012[+0.008,+0.093][+0.008, +0.093]+0.0360.104[−0.003,+0.079][-0.003, +0.079]

Both comparisons are significant on the 29 targets the base model does not reliably solve, with both intervals clear of zero, and neither is on the 21 it solves. The Best-of-N difference there is the do no harm result, and FK's apparent lead is not established either, its interval straddling zero.

That explains why the pooled FK comparison returns 0.104 where Best-of-N returns 0.012, despite the same pattern. FK is the stronger baseline at both ends, leading CAST less on the 29 and taking more back on the 21, so the same cancellation leaves a smaller margin. Averaging the two groups gives +0.051+0.051 against Best-of-N and +0.036+0.036 against FK, an effect 30% smaller on the same 50 targets, which is what moves the pooled p across the threshold (-0.003, 0.079). The pooled test therefore measures the composition of a difficulty balanced benchmark rather than the absence of an effect.

The mechanism behind that split is resampling. FK proposes from the unmodified base model and resamples on the reward, so it can only concentrate mass the base model already puts somewhere. Where the base model already places mass on the right site, concentrating it there is exactly right and FK wins, which is the solved row. Where it places almost none there is nothing for the reward to select, and concentrating harder makes it worse, which is the failed row. Resampling can only redistribute mass, and creating it where the base model puts none is what the gradient does, which is the transport the paper means. That is why we also have to trade-off between them.

W2 and Q2 together, dependence on map quality and performance against map accuracy.

Agreed that the map sets a ceiling, and here is the dependence measured rather than bounded by the oracle contrast. New table shows targets binned by distance from the true epitope to the nearest map candidate.

closest candidatenCASTBest-of-NgapCAST reachBest-of-N reach
under 6 Å320.4250.391+0.0342522
6 to 12 Å160.3020.203+0.099114
over 12 Å20.1110.164-0.05300

A better map is better, but an accurate map is not what the method needs. Accuracy helps monotonically, 0.425 under 6 Å against 0.302 at 6 to 12 Å, but the gain over Best-of-N runs the other way, +0.034+0.034 against +0.099+0.099, since where the map is right the base model often reaches the site unaided. The requirement is a candidate within about 12 Å, not a correct one, so the dependence is weaker than the oracle map ablation alone suggests. The outer bin holds only 2 targets, so we draw no conclusion from it.

As a control, Spearman rho between map error and outcome is −0.477-0.477 for CAST and −0.457-0.457 for Best-of-N. Best-of-N never sees the map, so its outcome is a map free measure of how hard the target is, and map error tracks it just as closely. Map error is therefore largely a proxy for target difficulty, and almost none of CAST's correlation is dependence on the map.

Suggested future work.

We agree, and an accurate epitope predictor is also an open problem.

Thank you again for your review. We have posted a detailed response and new measurements addressing the points you raised. As the discussion period is going to a close, we would be grateful to know whether our response resolves your concerns, and we are happy to provide anything further.

Thanks to the authors for these explanations. I already (weak) accept the paper, so I retain my score.

Thank you for reading our response. We are glad the concerns are addressed.

Your W1 and Q1 changed the reporting. The paired result on the 29 targets the base model does not reliably solve is now in the results section. The bootstrap intervals are in a new appendix table, for both Best-of-N and Feynman Kac. The claim is scoped to that regime in the text. That is what your Q1 asked.

Your W2 and Q2 added a table. It bins outcome by candidate map error. A Best-of-N control sits beside it, so a reader can see how much of the binning is target difficulty.

We also acted on your suggestion for future work. We ran an external epitope predictor over all 50 antigens. It now sits beside our own map in the candidate map table.

Thank you again for the care you gave the paper.

Reviewer_xXG1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors argue that co-folding models dock antibodies onto the wrong epitope, and that drawing more samples only deepens the mode the model already favours rather than relocating it, so naming a binding site and steering the sampler to it is what is needed. Conditioning-side restraints name a site but cost a trunk pass per candidate, while steering inside the diffusion loop diversifies the pose a complex adopts rather than the site where it binds. To this end, the authors propose CAST, a training-free framework that carries a binding site as an objective inside the diffusion loop. Specifically, a reference-free map of candidate sites is crossed with a sweep of arrival radii to form a grid of cells, and each particle in a single pool is assigned one cell and one MSA-subsampled conditioning latent for its whole trajectory, so that all conditionings share a single pool instead of needing one of their own; since the reward reads the paratope’s position alone, the tilt changes which site the antibody binds while leaving the pose at that site to the base model. Through experiments on 50 difficulty-balanced antibody–antigen targets and a second set of complexes that the unsteered model leaves unsolved, against an unsteered control and four steered baselines at a matched structure budget, the authors demonstrate that CAST brings more targets to an acceptable pose and raises mean max DockQ, with the gain concentrated on the targets the base model does not reliably solve.

Critical strengths
  1. The related-work section organizes prior steering methods along two axes, what they diversify and where they intervene, and the resulting grid makes the gap it claims concrete rather than sweeping: site-directed work has so far intervened at the model input, and diffusion-side work steers without naming a destination.

  2. Carrying a site hypothesis in the particle state is a natural way to explore many candidates at once, and pairing it with the conditioning latent is the better idea of the two, since it lets one pool reallocate trajectories across conditionings where separate pools would have fixed that allocation in advance.

Critical weaknesses
  1. While the authors honestly show what a wrong candidate site costs — A5 drops to the level of the unsteered control, and Figure 4 separates a map failure from a method failure — the ablations place the largest lever outside the method itself. An oracle map gains +0.111 and takes reach from 36 to 42, more than removing the gradient (−0.063) or shuffling the map (−0.054) costs. The paper argues that CAST takes a site as an argument so any epitope hypothesis plugs in, but that claim is never tested; a run substituting an external epitope predictor for the shipped heuristic map would show where the ceiling sits for a realistic hypothesis.
  2. FK is the only baseline sharing CAST’s reward, and so the one comparison that isolates the gradient, but it is run at a single arrival radius against CAST’s four — a difference the authors themselves invoke to explain FK’s advantage where the site is already found. The gap over all targets is not significant, and the significant result comes from the subset the unsteered control does not reliably solve. Richman et al., named in the related work as the closest method, runs the same twisted sampler under an RMSD potential and is discussed but never run, so the claim that a site-measurable reward beats a pose-moving one is argued rather than tested.
Questions for authors
  1. The authors note that ranking is a separate failure and that existing confidence estimates often cannot identify the best structure a pool contains, which is a fair reason to report pool-max. The metric nonetheless rewards the stratification CAST does and the baselines do not, since spreading a pool over many cells raises the maximum it contains while saying nothing about whether that structure can be found, and a user receives whatever the confidence head ranks first. Could the authors report the same comparisons at top-1, or how far down the ranking CAST’s best structure typically sits relative to BoN’s? If the advantage narrows there, that would not invalidate the coverage claim but would bound what the gain currently means in practice.
4 comments

We thank the reviewer for the sharpest question of the three, whether the map or the method does the work, and for identifying the amortisation as the better of its two ideas.

Clarification 1, pose moving rewards.

The closest runnable proxy for each half of ConforMix is already in the paper, and we failed to label them as such. ConforMix, Richman, Karaguesian, Suomivuori and Dror, NeurIPS 2025, runs the twisted sequential Monte Carlo sampler of Wu et al. 2023 under an RMSD potential that repels from the model's own default prediction. For the pose reward the proxy is AF3-ReD, Ohnuki and Okazaki, JACS Au 2026, a superposed Ca RMSD repulsion from the structures already produced, which reaches 0.311 against Best-of-N's 0.322 at the matched budget (Table 2). For the sampler it is A1, which replaces our joint pool with a separate particle set per conditioning, the arrangement it uses, at about 29 PFLOPs against CAST's 8.72 (Table 5), though A1 carries our site reward, not a pose one. The combined cell, a pose reward inside our joint pool, we have not run, and we would not expect it to close the gap where the base model fails. A pose reward names no site, ReD reaching 0.131 there against CAST's 0.182, while the pool itself moves the mean by 0.006 (A1). ConforMix itself costs about 3×3\times wall clock by its authors' own report, so a budget matched comparison would make it a 3×3\times compute arm.

W1, the oracle map as largest lever.

Arithmetic agreed, answered in the global response. Neither factor beats Best-of-N on the pooled mean or on borderline, so the oracle map pays only because the method converts it, and on that reading the ablation bounds what a better epitope predictor would buy.

W1, whether an arbitrary epitope hypothesis plugs in and an external epitope predictor.

A better map is better, but an accurate map is not what the method needs. The gain over Best-of-N is +0.034+0.034 where the closest map candidate is under 6 Å and +0.099+0.099 at 6 to 12 Å, with reach rising from 4 to 11, since where the map is already right the base model often reaches the site unaided. The binned table is in our reply to Reviewer isBc. The requirement is a candidate within about 12 Å, not a correct one.

We should also have been clearer that the map we ship is already a predicted map. It combines contact modes from an unsteered pilot pool with a farthest point tiling of the antigen surface, never seeing ground truth. Every CAST number is a result under a hypothesised epitope. We have run DiscoTope-3, Hoie et al., Frontiers in Immunology 2024, over all 50 CFA50 antigens at the thresholds of Table 4, the candidate map table (Appendix C.7), so the new rows sit beside ours. Both are capped at 15 candidates.

candidate map< 6 Å< 12 Å< 14 Åno usable map
Hybrid, ours, reference free3248490
DiscoTope-31536397 of 50

Our map reaches the true epitope within 6 Å on 32 targets against 15, and within 12 Å on 48 against 36, and produces a usable map on all 50 where DiscoTope-3 fails on 7. DiscoTope scores antigenicity with no antibody in view, so it cannot say which epitope is this one's. Ours is a shortlist, modes from the model's own placements of this antibody and tiles over the rest. The plug in claim therefore holds in the form the evidence supports, a predicted epitope map is useful to CAST as long as its error stays within about 12 Å.

W2, the Feynman Kac comparison.

The radius.

The radius matched comparison is in the paper. A3 is CAST run at a single arrival radius, so A3 against FK is like for like, and it gives 0.361 against 0.337 (Tables 2 and 5), CAST ahead by 0.024 of the pooled 0.036 gap. The umbrella sweep is a mechanism of CAST rather than a free parameter, so we do not give FK four radii.

The significance.

Correct, and decomposed in full in our isBc reply. The pooled p=0.104p = 0.104 comes from cancellation. CAST leads FK by +0.082+0.082 at p=0.003p = 0.003 on the 29 targets the base model does not reliably solve, and the two are indistinguishable on the 21 it does. That the significant result sits where the base model struggles is what the paper claims, not a limitation on it.

Q1, top-K pick and where the best structure sits in the ranking.

Both are in the global response. At top 1 the two arms are tied, 8 of 50 targets each, so the coverage gain does not reach a user who takes the head's first choice. It appears once a shortlist is inspected, 11 against 8 at K equal to 20, and CAST's best structure sits higher in the confidence ordering than Best-of-N's on every group. The revised discussion will say so.

Thank you again for your review. We have posted a detailed response and new measurements addressing the points you raised. As the discussion period is going to a close, we would be grateful to know whether our response resolves your concerns, and we are happy to provide anything further.

Thank you for the detailed response. My remaining concerns have been well addressed, and I will maintain my positive evaluation of the paper.

Thank you for reading our response. We are glad the concerns are addressed.

Your Q1 prompted the ranked selection experiment. It now closes the results section with a new appendix table. The top 1 result is stated plainly.

Your W2 asked about the significance. That decomposition is now in the appendix, for both Best-of-N and Feynman Kac. The radius-matched comparison is there as an explicit bound. A3 runs CAST at a single arrival radius against Feynman Kac. A reader can now see how much of the gap is the umbrella sweep.

Your W1 asked whether an external epitope hypothesis plugs in. We ran one over all 50 antigens and put it beside our own map. The claim is measured now, not argued.

Thank you again for your time during the discussion.

Reviewer_Af57OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CAST, a training-free sampling method for antibody-antigen cofolding that steers diffusion trajectories toward candidate epitopes. The method combines a reference-free candidate epitope map with gradient-based guidance and particle resampling across multiple MSA conditionings. Experiments on Protenix-v2 and OpenDDE show improvements in oracle best-of-pool DockQ and the number of targets for which an acceptable complex is sampled, particularly on difficult targets.

Critical strengths
  1. The paper addresses a highly relevant problem in antibody-antigen structure prediction, where scaling model compute at inference time could lead to repeated sampling incorrect binding sites. Separating site coverage from confidence-based ranking is a reasonable motivation.

  2. The experiments include several sampling baselines, compute-aware comparisons, ablations, and evaluation on a second cofolding backbone. The analysis of sampling breadth and performance on difficult targets is useful.

Critical weaknesses
  1. The main idea of using an epitope-dependent reward to guide diffusion sampling seems relatively straightforward and builds heavily on existing guided sampling and particle-based methods. The theoretical result on preserving the conditional pose distribution largely follows from the reward depending only on the binding-site coordinate, and does not directly establish the behavior of the approximate sampler. Overall, I find the methodological contribution incremental relative to its complexity.

  2. While CAST improves mean oracle DockQ from 0.322 to 0.373 over the default Best-of-N baseline, a stronger Best-of-N allocation reaches 0.366 (Table 3), cutting the lead to a much smaller one (although at greater computational cost). The improvement over Feynman-Kac steering is also not significant across the full CFA50 benchmark (Table 2). More importantly, removing resampling produces the same mean DockQ (0.373) and improves acceptable-pose coverage from 36 to 38 targets (Table 5). These results do not convincingly establish the necessity or benefit of the full proposed algorithm over simpler site-guided sampling approaches.

  3. The evaluation focuses entirely on oracle best-of-pool DockQ, which measures whether an acceptable structure exists among hundreds of generated samples rather than whether it can be identified in practice. While this is consistent with the paper's focus on site coverage, the practical benefit of the improvements remains unclear. The evaluation also emphasizes difficulty-selected targets, making it difficult to assess the broader usefulness of the method.

  4. The presentation is unnecessarily complicated for the central idea, and the main takeaways are obscured by extensive methodological details.

Questions for authors
  1. Given the results of Table 5, what is the practical advantage of the proposed particle resampling mechanism over simply applying gradient-based guidance toward candidate epitopes? Would the no-resampling variant perform similarly on OpenDDE and across different sampling budgets?

  2. How much of the improvement translates to structures selected using prediction confidence rather than oracle DockQ? Even a simple Top-k analysis would help establish whether the additional site coverage provides useful candidates.

  3. Could the authors compare against simpler epitope-conditioned sampling strategies using the same candidate map and computational budget? This would help distinguish the benefit of site-directed sampling from that of the proposed algorithm.

3 comments

We thank the reviewer for the most detailed engagement with our tables.

Clarification 1, how CFA50 was assembled.

CFA50 is difficulty balanced, 16 fails, 13 borderline and 21 the base model already solves, so 42% of the benchmark needs no site coverage help (Table 1, grouping in Appendix C).

Clarification 2, the compute column.

A stronger allocation reaches 0.366 and costs more. That 20×3620 \times 36 rung spends 15.40 PFLOPs against CAST's 8.72 at the same 720 structures, and the means are not separated, Wilcoxon p=0.77p = 0.77 (Section 4.1, Table 3). The extra conditionings buy no coverage either, reaching 1 of the 16 targets the base model fails against CAST's 4. On the collected 21-hard targets no allocation closes the gap at any budget, BoN's 2500 structure pool costing 5.3×5.3\times CAST's compute and still sitting 0.062 below (Figure 2c,d).

Clarification 3, what the resampling ablation removes.

A6 is not a simpler alternative. It is CAST with resampling switched off, keeping the site measurable reward, the map, the tilt and the umbrella sweep. Map alone reaches 0.310 and shuffled map plus gradient 0.319, both at or below BoN's 0.322, and on the failed group 0.127 and 0.167 without reaching CAST's 0.182 (Table 5), so what A6 retains is the contribution.

W1, novelty and Proposition 1.

What is new is what enters the particle state. Conditioning names a site but fixes it before the trunk, steering acts inside the loop but names no site, and does so either repulsively or from an oracle site. Ours is the only attractive site objective driven by a predicted map (Section 1). CAST carries both the site and the conditioning latent as particle state, so one pool sweeps S sites and K subsamples at N+KN + K rather than K×NK \times N, with no trunk pass per site, which Reviewer xXG1 calls the better of the two ideas and isBc novel and well motivated.

The other three pieces make that pool necessary and make it safe. The map is a prediction, so a grid of S candidates must be carried instead of one committed site. That grid is affordable only under the amortisation, and safe only because a site measurable tilt leaves the pose conditional untouched, which is Proposition 1, so a wrong candidate cannot damage a pose already right. On its scope, it states a property of the tilted target, and Appendix B.1 covers the sampler we run. The do no harm property is then measured, a null on the group BoN already solves.

W2 and Q1, why resample at all and FK (resampling method) concerns.

Proposition 1 holds whatever samples the tilted target, and weights and resampling are what tie the sampler to it at all, which pure gradient guidance lacks, targeting no distribution. We do not claim asymptotic exactness, which Appendix B.1 says our schedule forfeits by weighting only inside the window. Every number we report is a pool maximum, unaffected by that forfeit. Table 5 shows no pooled effect because the group effects cancel, not because the mechanism is inert. Resampling concentrates mass the reward already scores well, gaining 0.034 on borderline and 0.009 on solves, paying 0.039 on fails (Section 4.2). CAST therefore resamples early, taking the borderline and solved gains while leaving the failed group a pool that is still exploring. One may drop it, and the contribution holds either way, since with resampling off CAST still leads BoN by 0.051 pooled and 0.110 on fails, reach 38 against 26 (Table 5). The schedule is a trade we set, not a part the method rests on. On FK, the pooled p of 0.104 is the same cancellation, CAST ahead by 0.082 at p=0.003p = 0.003 on the 29 not reliably solved (isBc reply).

On OpenDDE the same cancellation appears, +0.020+0.020 against CAST on fails, the trade belonging to the schedule, not one model.

Q3, simpler epitope-conditioned baselines at the same map and budget.

The comparison is in the paper, unlabelled. Feynman Kac steering in Table 2 is exactly that baseline, the same reward and the same candidate map without the amortised joint pool, at 8.69 PFLOPs against CAST's 8.72, giving 0.337 and reach 30 against 0.373 and reach 36. A2, map with no gradient at 0.310, brackets it from below. Input level restraint conditioning is the one variant we cannot run, the Protenix-v2 checkpoint shipping no pocket embedder weights (Appendix A). A restraint commits to the candidate it encodes where the tilt only biases, which matters where the map is 6 to 12 Å off, the regime where committing hurts.

W3 and Q2, oracle only evaluation and confidence selection.

Addressed by New results 1 and 2 in global response.

W4, presentation.

Accepted for the ablation discussion, six arms in one paragraph with Table 5 in an appendix. We will bring that table into the main text and rewrite the paragraph around the three results that carry the method. On the central idea we disagree, see W1.

Thank you again for your review. We have posted a detailed response and new measurements addressing the points you raised. As the discussion period is going to a close, we would be grateful to know whether our response resolves your concerns, and we are happy to provide anything further.

As the discussion period is going to a close, we would be grateful to know whether our response resolves your concerns, and we are happy to provide anything further.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Global response

We thank all three reviewers. Reviewer isBc calls the approach novel and well motivated and Reviewer xXG1 calls the amortisation the better of its two ideas. Two questions recur across the reviews, whether the gain survives confidence based selection and how far it depends on map quality, and we answer both below with new measurements.

What the paper claims. CAST targets the structures a co-folding model does not already reach. It widens what the pool contains at a fixed budget, and Proposition 1 in Section 2.3 predicts that it does not disturb what the base model already gets right. We claim a sampler, not a verifier. What a pool does not contain cannot be ranked into it, so coverage comes first and selection second, and a better verifier has more to find in a CAST pool than in Best-of-N's. The two new results measure each in turn.

New result 1. Ranked selection

Af57 W3 and Q2, and xXG1 Q1, ask whether the gain survives confidence based selection. Selector is ranking_score, the base model's own confidence score. Counts are CFA50 targets reaching a CAPRI acceptable pose within the top K selected, at the matched compute budget. The groups and the oracle column are those of Table 1.

groupnarmK=1K=5K=20K=100K=360oracleconfidence rho
base model fails16CAST0000140.135
Best-of-N0000000.103
borderline13CAST11115120.182
Best-of-N1111460.015
base model solves21CAST79101518200.246
Best-of-N777815200.008
all 5050CAST810111624360.194
Best-of-N888919260.040

The last column is the Spearman correlation between confidence and true DockQ inside each pool, that is how rankable it is.

CAST overtakes Best-of-N from K=5K = 5 and the margin widens, 10 against 8, then 11 against 8, then 16 against 9.

CAST delivers a different gain in each regime, both visible here. On the 16 targets the base model fails it is coverage, where CAST places an acceptable structure into the pool on 4 targets while Best-of-N's pool holds none on any of the 16, at any K including oracle. On the 21 the base model already solves it is ranking, where both arms hold all 20 at oracle and CAST returns 15 of them inside the top 100 against Best-of-N's 8.

The ranking gain comes from rankability. Inside a CAST pool the confidence score correlates with true DockQ at 0.246 on that group against 0.008 inside Best-of-N's, so an off the shelf selector finds the good structures in one pool and not in the other.

On the failed targets, coverage comes first and ranking second. No selector can rescue a pool holding nothing acceptable, and CAST supplies an acceptable structure on 4 targets where Best-of-N supplies none. Ranking those 4 is a further step we do not claim, and New result 2 shows why, the base model's own confidence ranks the correct pose down.

New result 2. Why ranking stays hard on the failed targets

Ranking does not rescue the failed targets in either arm. Among the 4 targets CAST brings to an acceptable pose, none is ranked first. Selection is by ranking_score over all 50 targets at the matched compute budget, with both arms scored by the same code.

How the scores are computed. ranking_score is 0.8 ipTM +0.2+0.2 pTM, a TM score weighted by the model's predicted aligned error. We recompute it from the true aligned error, measured against the native structure, which gives what the model would have said had its error prediction been right. Both are confidence scores on the same 0 to 1 scale where higher is better.

grouparmnoracle percentileerr(pick)true gap
base model failsCAST1665%-0.015+0.061
Best-of-N1670%-0.066+0.029
base model solvesCAST2141%-0.043+0.287
Best-of-N2160%-0.001+0.249
all 50CAST5053%-0.045+0.218
Best-of-N5065%-0.038+0.171

Oracle percentile is where the pool's best structure by DockQ sits in the confidence ordering, averaged over targets, lower being better and 50% chance. err(pick) is true minus predicted score for the structure the head selects, so a negative value means the model overrates it. True gap is the true score of the oracle structure minus that of the selected one, so a positive value means a correct error prediction would have favoured the better structure.

On the failed group the head places the best structure worse than chance in both pools, 65% for CAST and 70% for Best-of-N against the 50% a random ordering gives, so it does not find it. Where the base model already solves, CAST's pool is rankable and Best-of-N's is not, 41% against 60. With the true aligned error the score ranks the best structure above the selected one on average in both arms, so the weakness is the predicted aligned error and not the pool. Best-of-N's selected structure is overrated by 0.066 there against CAST's 0.015, which we do not claim changes the top 1 result, 8 of 50 for both arms. Steering is not what breaks the head. Every CAST run also emits K minus 1 structures with no steering applied, which we drop from the scored pool. They are a within run control, the same model on the same target, and the model misjudges their aligned error by 28.8 Å on the failed group against 28.3 Å for the steered structures beside them. This matches the confidence weaknesses reported for antibody antigen complexes (Fromm, Ludaic and Elofsson, Bioinformatics 42(4) btag136, 2026, who also recompute their measures under perfect predicted aligned error as we do here).

CAST's contribution is putting an acceptable structure into the pool where Best-of-N has none. Selecting it needs a better verifier, a separate problem we state as a limitation and will add to the revised discussion.

On the oracle map ablation

Two reviewers read this as showing the largest lever sits outside the method. The arithmetic is right. What completes it is that the two factors are not separable. Map with no gradient, A2, reaches 0.310 and gradient on a shuffled map, A5, reaches 0.319, both at or below Best-of-N's pooled 0.322, where the two together give CAST 0.373 (Table 5, Appendix C.8). The oracle map, A4, pays +0.111+0.111 because the method converts map accuracy into pose accuracy, which makes it an upper bound on what a better epitope predictor would buy.

Measured directly against map error, the dependence is weaker than that bound suggests. The gain over Best-of-N is +0.034+0.034 where the closest candidate is under 6 Å and +0.099+0.099 at 6 to 12 Å, with reach 11 against 4, so the requirement is a candidate within about 12 Å, not a correct one. Map error also tracks Best-of-N's outcome almost as closely as CAST's, −0.457-0.457 against −0.477-0.477, and Best-of-N never sees the map, so map error is largely a proxy for target difficulty. Both results are in our reply to Reviewer isBc.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.