acceptodds
Withdrawn from ICLR 2027

Seed2Succeed: Test-Time Scaling Needs Breadth in Antibody–Antigen Structure Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Test-time scaling in antibody-antigen structure prediction offers a route to improve its accuracy, but its benefits depend on how prediction budgets are allocated. We present Seed2Succeed, a systematic study of seed breadth and sampling depth in AlphaFold3 using 113 FoldBench complexes and 2.26 million predicted structures. For each target, we compare 1,000 seeds 10 diffusion samples with 10 seeds 1,000 samples at a matched budget of 10,000 structures. We jointly analyze predicted epitope and paratope diversity, local structure diversity and global structure diversity. On 160 interfaces from 104 targets, broader seeding increases low-quality success rate from 43.1% to 60.0%, high-quality success rate from 16.3% to 25.0%. Broader seeding also expands contact coverage, while coordinate dispersion in the complementarity-determining regions remains nearly unchanged at full budget. However, additional sampling can reduce confidence-selected success, and full-pool oracle discovery numerically favors depth. Cluster analysis reveals structural populations with similar interface confidence but markedly different docking quality. These findings support seed breadth as a practical starting point on this benchmark and show that effective test-time scaling requires jointly considering candidate discovery, structural exploration, and ranking reliability.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AlphaFold 3, antibody–antigen complexes, random seeds, diffusion sampling, DockQ

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_HcpuOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates and compares the success rate of co-folding with varying depth (more samples in the diffusion step) and breadth (just more random seeds).

Critical strengths

Highly significant research questions that everyone wants answered. This comparison (broad vs. depth) requires substantial computing resources, and this paper did so and found meaningful implications for the community.

Critical weaknesses

The findings and conclusions are somewhat incremental. For result 4.1: We all already know that the DockQ score increases with more seeds (it's already well provided in the AF3 paper). For result 4.2, we all already know that more seeds/more samples will increase diversity (because folding is stochastic). The difference between broad and deep strategy in Fig 3 is not significant. For result 4.3, it's again obvious that the success rate increases with more diverse co-folding structures and poses.

This paper's strongest claim would be "full-cohort oracle discovery numerically favored depth" or " additional sampling can reduce confidence-selected success, and full-pool oracle discovery numerically favors depth". This claim is not well supported by experiments. The reported 95% interval includes zero. This does not establish that deep strategy is better than broad.

Questions for authors

After this huge investigation, the findings somewhat reconfirm what we already know. Is there any strategy to combine both breadth and depth and identify the best sweet spot that increases the success rate within the same budget?

Reviewer_tsbaOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates test-time scaling for antibody-antigen structure prediction using AlphaFold 3, exploring how inference budgets should be allocated between the number of random seeds and diffusion samples per seed. The authors evaluate 113 complexes from FoldBench and generate 2.26 million predicted structures, mainly comparing two sampling configurations: BROAD (1,000 seeds × 10 samples) and DEEP (10 seeds × 1,000 samples). Prediction accuracy is assessed using confidence-selected DockQ, oracle discovery, and several structural diversity metrics. The main finding of the paper is that broader seed sampling improves confidence-selected prediction quality, although deeper sampling achieves numerically higher oracle success at the full budget.

Critical strengths
  1. The authors conduct a large-scale empirical study involving over 2.26 million predicted structures across 113 antibody–antigen complexes from FoldBench. The experiments systematically compare different allocations of random seeds and diffusion samples under matched structure-count budgets, providing a useful empirical characterization of inference-time sampling behavior in AlphaFold 3.

  2. Beyond reporting prediction accuracy, the paper examines several complementary aspects of test-time scaling, including epitope/paratope contact coverage, structural diversity, and confidence-based ranking. The analysis also highlights potential limitations of existing confidence metrics in distinguishing correct from incorrect docking poses.

Critical weaknesses
  1. The paper's central observation that broader seed sampling can improve antibody-antigen structure prediction is not particularly novel. The original AlphaFold 3 paper (Abramson et al., 2024, Fig. 5a) already evaluated inference-time scaling with up to 1,000 seeds × 5 diffusion samples and explicitly reported that increasing the number of seeds was more beneficial than increasing diffusion samples per seed. More recently, Protenix-v2 (Zhang et al., 2026, Fig. 3) also systematically evaluated inference-time scaling up to 1,000 seeds, showing that improved model quality can substantially reduce the required sampling budget. Given these prior findings, it is unclear what substantially new scientific insight this paper provides beyond existing practices.

  2. The main experiments compare 1,000 seeds × 10 samples against 10 seeds × 1,000 samples, requiring 10,000 structure predictions per target. Such an inference budget is unlikely to be practical for antibody design campaigns, where hundreds or thousands of candidate binders may need to be evaluated and filtered. Although the paper includes smaller-budget experiments, the improvements in these settings are considerably less compelling. For instance, at a budget of 100 structures, the acceptable-quality Top-1 success rate increases only from 54.4% (10 seeds × 10 samples) to 56.9% (100 seeds × 1 sample). Furthermore, matching the number of generated structures does not necessarily imply equivalent computational costs, and the paper does not really establish an optimal sampling allocation under realistic GPU-time constraints.

  3. Despite the substantial computational resources invested, the paper does not introduce a new sampling method or establish a clear recommendation beyond the existing practice of multi-seed sampling and confidence-based ranking. Moreover, the conclusions depend on the evaluation criterion: while BROAD improves confidence-selected success, DEEP achieves numerically higher full-budget oracle discovery (90.3% versus 87.6%). Thus, the results do not demonstrate that broader sampling is universally preferable, nor do they provide a principled strategy for choosing between seed breadth and sampling depth. Overall, I find the contribution more comparable to a descriptive empirical study or technical report than a sufficiently substantial research contribution for a venue like ICLR.

Questions for authors
  1. Given that Abramson et al. (2024) already reported that increasing the number of seeds is more beneficial than increasing diffusion samples per seed for antibody–antigen prediction, what do the authors consider as the main novel scientific insight of this study?

  2. What sampling configuration would the authors recommend for practical antibody design campaigns, where the inference budget per candidate is limited? Would the conclusions remain the same when comparing sampling strategies under matched GPU-hours rather than matched numbers of structures?

  3. How do the authors reconcile BROAD's improvement in confidence-selected success with DEEP's numerically higher oracle discovery at the full budget? To what extent does the advantage of BROAD reflect limitations in confidence-based ranking rather than improvements in candidate generation?

[1] Abramson, J., et al. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature, 630, 493–500. Figure 5a.

[2] Zhang, Y., et al. (2026). Protenix-v2: Broadening the Reach of Structure Prediction and Biomolecular Design. bioRxiv. Figure 3

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

The paper asks how a fixed inference budget should be divided between random seeds and diffusion samples in AlphaFold 3 for antibody-antigen prediction. On the 113 antibody-antigen targets of FoldBench, the authors generate two ensembles per target at a matched budget of 10,000 structures -- BROAD (1,000 seeds x 10 samples) and DEEP (10 x 1,000) -- for 2.26M structures in total. Intermediate configurations are prefixes of these two ensembles rather than independent runs.

Evaluation runs on two tracks. Selected success is per interface, on a fixed roster of 160 interfaces from 104 targets, taking the best DockQ among confidence-ranked Top-k candidates. Oracle discovery and diversity are per target over all 113, using a bottleneck DockQ that requires one model to be accurate at every interface at once. Diversity is measured by contact coverage, maximum pairwise antibody RMSD and full-CDR RMSF in an antigen-fixed frame, plus DBSCAN clustering.

The main results are that BROAD improves Top-1 by 8.75 points at DockQ > 0.49 and > 0.80; that adding samples at fixed seed count can lower Top-1 even though the candidate sets are nested; that oracle discovery mildly favors DEEP; and that in two case studies mean ipTM barely separates a large low-quality cluster from a smaller high-quality one.

Critical strengths

I think the question is a good one and it is asked correctly. How to spend an inference budget is something people actually decide every day in this setting, and it is usually decided by folklore. Separating seeds from samples is the right decomposition, because they are not interchangeable: a new seed resamples the conditioning, while samples within a seed share the whole trunk computation. Holding S x D fixed and sweeping the split is a well-posed experiment, and running it at 2.26M structures is a real commitment. I would like a careful version of this paper to exist.

The nested-prefix argument is the sharpest thing in the paper. Since all configurations are prefixes of one ensemble, oracle best DockQ cannot decrease, so the drop in Top-1 at fixed seeds (96 -> 93 -> 91 at 1,000 seeds) must come from ranking displacing a better structure. That is a clean separation of selection failure from generation failure, and it is what makes the Top-1 / oracle divergence interpretable instead of merely confusing.

Critical weaknesses
  1. My main concern is that the paper names ranking as the bottleneck and then does not study ranking at all. The evidence is strong and it is the authors' own: Top-1 sits around 52-57% while oracle is around 88-90%; adding candidates to a nested set lowers Top-1; and in Table 6 mean ipTM does not order the clusters, with 8QZ3/BROAD actually inverted (0.5235 vs 0.5216 for mean DockQ 0.024 vs 0.777). After establishing this, the conclusion is that scaling should jointly address generation and ranking, which restates the problem. I would expect at minimum a ranking_score vs DockQ scatter and calibration curve over the 2.26M structures -- selection uses ranking_score throughout, yet Figure 5 shows ipTM -- and some attempt at cluster-based selection, since DBSCAN has already been run on all 226 tasks and Table 6 is itself an argument that majority rules should fail. Even four rerankers that all fail would be a contribution.

The same gap sits one level deeper. The whole story rests on seeds perturbing the upstream representation while samples only perturb decoding, but this is never measured. Section 4.2 says the data do not identify the responsible upstream component and then drops it. Looking at how much the pair representation, or the antibody-antigen cross-block distogram, actually moves across seeds is cheap next to generating 2.26M structures, and it would close the chain.

  1. My second concern is that no usable recommendation comes out, and I suspect a single global recommendation is the wrong target anyway. The allocation answer is in the data but never stated: rearranged by fixed budget, 1000x1 (0.600) > 100x10 (0.575) > 10x100 (0.544) > 1x1000 (0.469) at B = 1,000, and the same ordering holds at B = 100 and B = 10. That supports a simple rule -- spend on seeds, set D = 1 -- and it also shows that the recommended 1000x10 is not the optimum of the authors' own grid, since 1000x1 reaches 96/160 with a tenth of the structures and 100x1 already matches 1000x10. There is no budget-to- allocation table, no saturation point and no stopping rule, though the data would support all three. Separately, the comparison is normalized by structure count rather than compute: 1,000 seeds means 1,000 full trunk passes, 10 seeds amortize the trunk over 1,000 samples, and Appendix A concedes that the recycle and diffusion-step counts are not recoverable.

3.The last is the heterogeneity the paper reports and then averages away. BROAD wins antigen coverage on 63 targets and loses on 31; native CDR RMSD favors BROAD on 64 and DEEP on 49; in Table 6 the good population is larger under BROAD for 8QZ3 but under DEEP for 8TS0. These are presented as caveats against universal dominance. I read them as saying the question should be when breadth is worth it, not whether. Since seed noise enters largely through MSA subsampling, across-seed conditioning dispersion should track antigen MSA depth, and antigen familiarity probably matters too -- FoldBench is clean for AF3 by construction, but chain-level identity is a weak notion of novelty for antibodies, and independent co-folding work reports a 38-point gap between recurrent and novel antigens. Stratifying Tables 2 and 4 this way is a day of work on data already generated, and I think it is the shortest path from the current descriptive comparison to something actionable.

Questions for authors
  1. What does ranking_score vs DockQ look like over the 2.26M structures, as a scatter and a calibration curve, per target?

  2. Did you try any ranking experiments? Using the existing DBSCAN output, how do simple cluster-based rules (largest cluster, second-largest, centroid consensus, diversity-aware reranking) compare with ranking_score Top-1?

  3. Is the oracle-minus-Top-1 gap different between BROAD and DEEP?

  4. Have you looked at how much the upstream representation changes when the seed changes, and how that varies across targets? Does it track antigen MSA depth, and does it predict whether breadth wins on that target?

  5. If Tables 2 and 4 are stratified by MSA depth or by antigen recurrence, do the conclusions hold in both strata?

  6. Why recommend 1000x10 when 1000x1 reaches 96/160 at a tenth of the budget and 100x1 already matches 1000x10? Can you state a budget-to-allocation rule and a saturation point?

  7. Section 3.1 says execution time was recorded for every task, but the Limitations report timing for only 25 paired targets. Is the full timing available? Together with recycle and diffusion-step counts, this would let the comparison be normalized by compute.

  8. Section 4.1 compares a Top-25 rate over 1,000 structures (0.688) against Protenix-v2 (0.65) and OpenDDE (0.70), which are Top-1, on a different denominator. Both baselines also have training cutoffs later than FoldBench's reference date while AF3 does not. How should this be read? I would suggest removing it or restating it at matched Top-1.

  9. How does this relate to Section 1.3 of the OpenDDE report, which does seed-dimension test-time scaling on the same benchmark and reaches a similar conclusion about ranking? It is cited but only for numbers.

  10. Will the 2.26M structures be released? Appendix D suggests only the summary bundle accompanies the paper, but the dataset is the first contribution.

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

Related papers

Discussion (0)

Sign in to comment.