acceptodds
Under review as a conference paper at ICLR 2027

AbPilot: Steering Diffusion Co-folding Models for Antibody Design via End-to-End Differentiation

Abstract

Modern all-atom co-folding models have sharply improved antibody–antigen structure prediction, yet their stochastic diffusion structure decoders make direct gradient-based antibody design difficult. Existing methods therefore either rely on simpler predictors or bypass the diffusion process, underexploiting the full capacity of co-folding models. We present AbPilot, which enables end-to-end differentiation through the native diffusion sampling process and propagates atomic-level structural and confidence objectives back to antibody sequence for steered generation. The gradient reaches the coordinates the model actually emits, so the interface can be controlled at atomic resolution. Using adjoint-state recursion, AbPilot reduces activation memory to O(1) with respect to diffusion depth, and further introduces orthogonal rectification to mitigate predictor hacking using an external structural verification as a search signal. On a new 12-target de novo antibody design benchmark, AbPilot achieves in silico success rates by more than an order of magnitude over the strongest baseline for both VHH and scFv design. Finally, prospective VHH design against mICAM-1, a target with no previously reported VHH binder, yields nanomolar binders, demonstrating that the proposed end-to-end differentiation can turn modern diffusion co-folding models into effective antibody design engines.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
End-to-End Differentiation, Antibody Design, Co-folding Models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_YkJCOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

De novo antibody design by inverting a co-folding model: optimise CDR sequence logits so the model’s own structural and confidence outputs report a favourable complex, as BindCraft does with AlphaFold-Multimer. Modern all-atom models like Boltz-2) predict antibody–antigen complexes better but decode with a diffusion sampler, and unrolled backpropagation through T denoising steps costs O(T) activation memory. Existing work avoids the sampler leaving objectives defined on emitted coordinates unavailable. The authors propose a solution ABPilot . It differentiates the sampler with a discrete adjoint recursion and trajectory replay. The forward pass runs graph-free while caching coordinates and per-step noise, and the backward pass reconstructs one transition at a time. Each step at t is rebuilt from its cached coordinates and noise, differentiated locally, then discarded. The authors claim that this yields the first end-to-end gradient through a diffusion co-folding model at activation memory constant in depth, that the resulting coordinate gradients enable geometric objectives unavailable to trunk-level methods, and that periodic re-ranking by an independent predictor limits model-specific reward hacking.

Critical strengths
  1. The paper is well written and well motivated
  2. The ablation against a distogram surrogate changes exactly one thing, whether the geometric gradient starts in sampled coordinates or in a predicted distance distribution, and finds that the surrogate reaches lower confidence loss but worse downstream acceptance (Appendix H Table 5), which helps generalise past antibodies to ither predictor-inversion methods 3.The authors provide a practical insight into what is feasible. Unrolling fails at T = 2 on a real job while the adjoint holds constant to T = 200 (appendix B), and the benchmark gap over four baselines on 11–12 of 12 targets is too large to be run-to-run noise (Appendix I).
Critical weaknesses

AlignProp (Prabhudesai et al., 2024) is not cited, and the claim to end-to-end differentiation rests on its absence. AlignProp backpropagates the reward gradient through the denoising chain of a diffusion model and makes it fit in memory with gradient checkpointing, reporting a reduction from 512 GB to 15 GB on a single GPU, which is a similar problem, and solution as section3.2 and Fig. 4.

Questions for authors

Please position the work against AlignProp (Prabhudesai et al., 2024), which backpropagates through the full denoising chain with gradient checkpointing. This has direct implication on appendix A Specifically: how does the adjoint recursion of section 3.2 differ from checkpointing at timestep granularity, beyond the replay of ε_t required by AF3’s rigid-body augmentation, and does any part of the memory result not follow from checkpointing alone? If the difference is one of application rather than procedure, please say so and identify which contributions you would have the reader weigh instead.

6 comments

We thank the reviewer for bringing AlignProp [1] to our attention.

Methodological differences. As AlignProp was originally developed for text-to-image diffusion models, we adapt its official implementation (https://github.com/mihirp1998/AlignProp/) to antibody design for a fair comparison, using the same frozen co-folding predictor, sequence optimization variables, and molecular objectives as AbPilot.

Both methods support gradient propagation through diffusion trajectories using recomputation. However, our discrete adjoint formulation explicitly separates forward sampling from backward differentiation , computing local vector–Jacobian products (VJPs) at each denoising step and accumulating gradients with respect to the shared conditioning before propagating them to sequence logits.

Importantly, unlike Gaussian truncation in AlignProp, which stops gradients through network predictions at selected denoising steps, our method preserves all diffusion-step gradient pathways . By explicitly replaying the stochastic variables from the forward trajectory, we recover the full discrete gradient, subject to the predictor's existing stop-gradient operations, without introducing additional temporal truncation.

Gradient fidelity evaluation. To quantify this distinction, we evaluated a 438-residue DerF7–scFv complex using 25 diffusion steps, three paired random seeds, and five Gaussian cutoff samples per initial input.

We compared each computed gradient g^\hat{\mathbf{g}} against the untruncated reference gradient g\mathbf{g} under identical inputs and stochastic trajectories, using three metrics:

Relative gradient error:

∥g^−g∥2∥g∥2\frac{\|\hat{\mathbf{g}}-\mathbf{g}\|_2}{\|\mathbf{g}\|_2}

Cosine similarity:

g^⋅g∥g^∥2∥g∥2\frac{\hat{\mathbf{g}}\cdot\mathbf{g}}{\|\hat{\mathbf{g}}\|_2\|\mathbf{g}\|_2}

Angular deviation:

\frac{\hat{\mathbf{g}}\cdot\mathbf{g}}{\|\hat{\mathbf{g}}\|_2\|\mathbf{g}\|_2} \right)$$ | Metric | AbPilot (local VJP) | Gaussian-truncated AlignProp | | ------------------------------ | ------------------------ | ---------------------------- | | Median relative gradient error | **6.98 × 10⁻⁷** | 98.89% | | Relative error range | 6.43–7.04 × 10⁻⁷ | 77.80–99.74% | | Median cosine similarity | **≈ 1.0** | 0.1576 | | Median angular deviation | **≈ 0.000040°** | 80.93° | **Our formulation reproduces the untruncated reference gradient with near-machine-precision agreement** , preserving both its magnitude and direction. In contrast, Gaussian truncation introduces substantial gradient discrepancies, with angular deviations ranging from 50.22° to 85.45°, indicating changes in gradient direction rather than merely magnitude. **Memory efficiency.** Our approach manages memory by constructing and releasing local computation graphs during the backward sweep, without discarding diffusion-step gradient pathways. Across three seeds and 25 consecutive optimization updates per seed, our encoder-recomputation variant achieves **approximately 7.1% lower mean peak allocated memory** than the AlignProp implementation. **Overall, these results demonstrate that AbPilot preserves full-gradient fidelity while achieving modest additional memory savings in the evaluated setting.** We emphasize that our contribution is not a fundamentally different mathematical gradient from exact checkpointed backpropagation, but an explicit trajectory-level differentiation formulation integrated into antibody co-folding optimization. This enables direct sequence optimization against structure-dependent geometric objectives while maintaining precise control over gradient propagation. [1] Prabhudesai, Mihir, et al. "Aligning text-to-image diffusion models with reward backpropagation." (2023).

Dear Reviewer,

We would appreciate it if you could take a look at our rebuttal and let us know whether your concerns have been adequately addressed. As the discussion period is approaching its end, we would particularly appreciate any further feedback you may have.

We would be happy to clarify any remaining questions.

Best regards,
The Authors

Dear Reviewer,

As today is the final day of the discussion period, we kindly ask if you could review our rebuttal and let us know whether your concerns have been addressed.

We have invested significant time and effort into carefully addressing your questions, including conducting additional experiments and providing detailed clarifications. We would sincerely appreciate your feedback.

Thank you for your time and consideration.

Best regards, The Authors

Reviewer_fJnyOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

AbPilot optimizes antibody CDR sequence logits through a frozen diffusion co-folding model. It combines discrete adjoint differentiation, coordinate and confidence objectives, and optional orthogonal rectification. The paper evaluates VHH and scFv design on 12 targets and reports VHH wet-lab validation.

Critical strengths
  1. The described method is interesting even though it is not novel.
  2. The reported results are promising. Ablation shows that diffusion differentiation achieves high averaged AF3 success against e.g. the distogram surrogate.
  3. Wet lab experiments is a strength. 6 of 50 selected VHH designs pass the ELISA threshold.
Critical weaknesses
  1. The distinct technical contribution is unclear. For example, Mosaic (https://github.com/escalante-bio/mosaic/), through Joltz (https://github.com/nboyd/joltz/), already exposes sampled coordinate objectives through a Boltz-2 diffusion sampler with JAX checkpointing. This overlaps with the central contribution of this work. The AbPilot's useful differences need to be distinguished from existing differentiable sampling.
  2. The implementation cannot be adequately checked. No code or complete configuration is supplied or linked in the pdf.
  3. The benchmark does not directly test final epitope-specific success. The ipTM, pLDDT, and scRMSD criteria omit explicit epitope attraction and framework orientation checks. Figure 2 reports epitope and framework success during optimization, but not for the final AF3 structures.
  4. The ablation and efficiency comparisons need clarification. Table 5 does not specify its rectification model. The scope of the Boltz-2 filter in Appendix H is unclear. AF3 rectification uses the final evaluator for candidate selection. Seems, like only BoltzGen's top 1000 candidates per target are evaluated, and costs within the 24-hour budget are not broken down by stage.
  5. Wet-lab validation covers one target and reports affinity for one design. Replicate counts, uncertainty, ELISA negative-control identity, and mICAM-1 construct residue ranges are unspecified. Also the experimental success rate is misleading since it is calculated from the designs that passed in silico threshold, but not all designs.

An antibody-specific empirical contribution could justify acceptance without a new differentiation mechanism. And the benchmark results as well as the reported wet-lab experiments are promising. However, multiple uncertainties in the manuscript coupled with the lack of released implementation limit the evidence. The implemented objectives, complete protocol, and analysis of existing structures could change my assessment.

Questions for authors
  1. What specifically differs AbPilot from Mosaic/Joltz? A technical explanation would clarify novelty and contribution.
  2. Eq. 14 divides an unsquared distance by a squared threshold, inconsistent with the stated 12 Angs. threshold in table 4. Can you clarify?
  3. What fraction of AF3-accepted designs satisfy the epitope attraction and "framework-away constraints"? Analysis of existing structures would test epitope-specific success.
  4. Which rectification model and filters are used for each Table 5 method? Does base AbPilot use Boltz-2 candidate selection as in Appendix F? Please give rate denominators, stage counts, and costs.
  5. Please define the ELISA negative control, design/assay mICAM-1 constructs, replicate counts, and available ELISA/BLI uncertainty.
6 comments

We thank the reviewer for their thorough and constructive feedback. We address each point below.

W1 & Q1: Technical Differences from Mosaic

We clarify the methodological differences with Mosaic [1] and provide additional comparisons to better establish AbPilot's contributions.

Differentiation through diffusion. In Mosaic's default binder-design configuration, the binder–target contact objective is computed from distogram predictions derived directly from trunk representations, rather than sampled coordinates. Consequently, gradients from this objective bypass the diffusion sampler. The default sequence-recovery objective also stops gradients through its structure-dependent sequence targets.

For objectives requiring differentiation through diffusion, Mosaic employs JAX gradient checkpointing (jax.checkpoint) to recompute intermediate activations during backpropagation. While both checkpointing and our discrete adjoint formulation reduce memory consumption through recomputation, our approach explicitly decouples forward sampling from backward differentiation. Please refer to Q2 of our global response for our detailed response about the difference between ours and checkpoint.

More importantly, our contribution extends beyond memory-efficient differentiation. AbPilot integrates trajectory-level differentiation into an end-to-end antibody optimization framework, enabling direct optimization of structure-dependent geometric objectives through diffusion co-folding models rather than relying solely on distogram-based surrogates.

Additional comparisons with Mosaic. To further validate our approach, we added two Mosaic baselines to Table 1, and set the base model as Boltz-2 :

  • Mosaic (Default): Adopts Mosaic's original design objectives and loss weights, adapted to our benchmark tasks, following its official implementation https://github.com/escalante-bio/mosaic/blob/main/examples/boltz_notebook.py.
  • Mosaic (Matched): Uses the same optimization objectives and loss weights as AbPilot while retaining all other settings from Mosaic (Default), enabling a controlled comparison involving diffusion-dependent gradients.

AbPilot substantially outperforms both variants (Table 1, Figure 10 and 12). It achieves success rates of 28.4% for VHH and 28.8% for scFv, compared with 4.1% and 4.7% for Mosaic (Default), respectively. Mosaic (Matched) achieves 0% success, potentially due to conflicts between diffusion and confidence gradients, which AbPilot explicitly separates for finer control (Figure 7f). These results demonstrate that AbPilot's performance gains cannot be explained solely by differences in optimization objectives, supporting the effectiveness of our integrated optimization framework.

MethodVHHscFv
AbPilot + AF3 rect.28.4%28.8%
AbPilot21.6%26.0%
AbPilot + PX rect.26.1%25.3%
Mosaic (Default)4.1%4.7%
Mosaic (Matched)0.0%0.0%

W2: Implementation and Reproducibility

We have added detailed pseudocode and implementation details to the appendix to improve reproducibility. We also plan to release the complete implementation shortly after the ICLR decisions are announced.

W3 & Q3: Epitope-Specific Geometric Success

To directly assess epitope-specific design success, we conducted an additional geometric analysis of both the original design structures and their corresponding AF3 predictions.

For the specified 10-target epitope cohort, we evaluated designs passing the computational success criteria. Geometric success was defined by simultaneously satisfying an epitope distance ≤ 10 Å and a framework distance ≥ 15 Å.

MethodDesign structuresAF3-MSA predictions
Diffusion differentiation (Ours)59.88%73.85%
Shortcut21.30%21.73%
Distogram surrogate42.86%45.51%

AbPilot consistently achieves the highest geometric success rates, outperforming both shortcut differentiation and distogram-based optimization. Notably, this advantage persists after AF3 re-prediction. These results provide direct evidence that AbPilot achieves more effective epitope-specific geometric control beyond its original design structures.

[1] Boyd, N., Guns, S. & Escalante Bio. Mosaic. https://github.com/escalante-bio/mosaic (2025).

W4 & Q4: Rectification, Candidate Selection, and Evaluation Protocol

We expanded revised Section 3.4 to explicitly describe the rectification model, candidate filtering, and evaluation protocol used in Table 5 (now Table 6).

In-loop rectification and candidate selection. All rectification-enabled configurations in Table 5 use AlphaFold 3 (AF3) as the rectification model. During iterative policy refinement (Section 3.4), AbPilot generates discrete sequences along with Boltz-2-predicted structures and confidence metrics. Candidates passing the initial Boltz-2 screening criteria (ipTM ≥ 0.7, CDR pLDDT ≥ 0.7, and ipAE ≤ 15 Å) are further evaluated using AF3 to guide candidate selection and subsequent optimization rounds to reduce the computational cost of in-loop evaluation. Importantly, this AF3 evaluation is performed internally during the design process for rectification and is distinct from the independent final evaluation used to calculate success rates.

Final success-rate evaluation. After the design process is complete, candidates passing the Boltz-2 screening criteria undergo a separate AF3 evaluation under the common success criteria. Specifically, the denominator is the number of Boltz-2-filtered candidates, while the numerator is the number satisfying the final AF3 success criteria, consistent with Table 1.

Base AbPilot does not use AF3 for in-loop rectification, relying instead on Boltz-2 predictions for candidate selection (Appendix F). However, it follows the same independent final AF3 evaluation protocol as the other methods. More detailed explanations are provided in Q1 of the new response to Reviewer CYuq

BoltzGen evaluation. Due to its substantially larger candidate pool (nearly 4,000 candidates per target, approximately 50,000 candidates in total) and computational constraints, we evaluate the top 1,000 BoltzGen candidates per target, selected using its official filtering and ranking procedure (https://github.com/HannesStark/boltzgen/blob/main/src/boltzgen/task/filter/filter.py). This shortlist is then evaluated under our common success criteria. The corresponding success-rate numerators and denominators are reported in Tables 1 and 6.

The average computational cost is distributed as follows:

StageGPU-hoursShare
Candidate generation12.1850.8%
Boltz-based screening and filtering2.5910.8%
In-loop AF3 rectification9.1037.9%
Initialization and other overhead0.120.5%
Total design budget24.00100%
Final AF3 evaluation with MSAs (excluded)8.60—
Total including final evaluation32.60—

W5 & Q5: Experimental Validation and Reporting

Experimental success rate. Consistent with the reporting protocols of BindCraft [2] and Germinal [3], our experimental success rate is calculated among designs that passed computational filtering and proceeded to experimental validation, rather than all generated candidates.

ELISA negative control. We used an unrelated FLAG-tagged nanobody targeting influenza hemagglutinin (HA) as the negative control. This control showed an OD405 of 0.077, whereas the mICAM-1-binding designs produced signals more than threefold higher, supporting target-dependent binding in the assay.

mICAM-1 construct. The construct comprises residues 28–115 of mouse ICAM-1 (UniProt P13597), corresponding to the D1 Ig-like domain. It was cloned into a pET-28a vector with a C-terminal 6×His tag, expressed in E. coli, and purified using Ni-NTA affinity chromatography. Protein concentration was determined by absorbance at 280 nm.

Experimental replicates and uncertainty. ELISA measurements were performed in two independent experiments. No technical replicates were included within individual experiments; therefore, intra-assay standard deviations are unavailable.

BLI measurements were performed using AR2G biosensors, yielding a binding affinity of K_D = 65.8 ± 1.6 nM, with a global-fit R² = 0.97.

We incorporated these experimental details in the revised Appendix L.

Q2: Correction of Equation 14

We thank the reviewer for identifying this inconsistency. The squared denominator in Eq. (14) is a typographical error. In our implementation, the mean Euclidean distance is normalized by the predefined distance threshold of 12 Å, without squaring. We will correct Eq. (14) accordingly.

[2] Pacesa, Martin, et al. Nature 646.8084 (2025): 483-492.

[3] Mille-Fragoso, Luis S., et al. Nature biotechnology (2026): 1-10.

Dear Reviewer,

We would appreciate it if you could take a look at our rebuttal and let us know whether your concerns have been adequately addressed. As the discussion period is approaching its end, we would particularly appreciate any further feedback you may have.

We would be happy to clarify any remaining questions.

Best regards,
The Authors

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

My take on the paper is that it is an interesting implementation of gradient checkpointing approach for antibody design with a considerable strong take on the implementation including engineering the approach to fit a nominal GPU. This makes it a genuine contribution to the ICLR venue taking it from an implementation specific angle, the derivation of the adjoint recursion and how it is applied to antibody design. The wet lab component is a strong addition to thi paper.

Critical strengths

The problem is well chosen and the framing is appropriate. The observation driving the paper is accurate: AFM’s deterministic decoder is what made BindCraft-style design possible, and the shift to diffusion decoders in AF3-class models traded that differentiability for accuracy. Most of the field responded by retreating to weaker predictors or to trunk-level surrogates. Asking whether the trajectory itself can be differentiated is appropriate in my mind, and the paper states the trade-off clearly.

The distogram surrogate represents the best piece of experimental design in the paper. Rather than comparing against BoltzDesign1 — which would have confounded the comparison with a different base model and different confidence terms, they re-express their own ℓ_epi and ℓ_FW-away on the trunk distogram, which I think adds value to cross-model comparisons.

The benchmark also addresses a real gap in the subfield. Antibody design papers have been evaluated on handful-of-target sets. The paper presents twelve targets drawn from established binder-design literature, both epitope-constrained and unconstrained, two frameworks, matched compute, and an evaluation criterion taken from prior work rather than new evaluations, which makes it easy to follow and (re)implement.

Critical weaknesses

The authors don't explain what all-atom layout the diffusion module sees when fed soft sequence probabilities. This is somewhat confusing in terms of reading what the paper is saying and what has been implemented.

The authors have not presented any error bars/ repeated runs on the data presented (table 1), and the wet-lab presentation does not present any comparisons to other design methods and no epitope-competition control.

Questions for authors
  1. What atoms does the diffusion module operate on during inner-loop optimization? When the input is a soft mixture over residue types, each with a different number of side-chain atoms, what atom layout does D_θ actually see — and how is ∂L_x/∂z meaningful when the atoms in x_0 don’t correspond to the residues z is optimizing?

  2. Does Boltz-2 at T = 25 produce reliable structures? The method only works in a narrow window around this depth, but the model ships at T = 200. Can you show, on held-out antibody–antigen complexes with known structures, that 25-step sampling doesn’t degrade prediction quality — and if it does, what is the optimizer actually fitting to?

  3. How does this differ from applying torch.utils.checkpoint to each denoising step with preserve_rng_state=True? If the answer is that it doesn’t, the contribution is the application and perhaps the authors should explicitly say so.

  4. Can you reproduce at least one baseline’s published success rate under its own protocol, then show what the matched-budget does to that number? Without this, the 8–17× headline could reflect the evaluation harness as much as the method.

5 comments

We thank the reviewer for their thorough and constructive feedback. We address each point below.

W1 & Q1: Atom Layout and Relaxed-to-Discrete Optimization

We have added a dedicated analysis of the relaxed-to-discrete optimization gap in the revised Appendix E , examining optimization trajectories and discrepancies between soft-sequence predictions and discretized candidates. Please refer to Q1 of our global response for our detailed response.

W2: Statistical Uncertainty and Experimental Validation

Statistical uncertainty. Figure 2 already reports error bars across repeated runs, providing evidence of the robustness of our optimization framework. While Table 1 does not report error bars, the substantial performance improvements provide strong evidence of our method's effectiveness. Repeating the full benchmark with additional seeds would require considerable computational resources and is not feasible within the rebuttal period.

Experimental validation. Our wet-lab experiments provide preliminary evidence that AbPilot-generated designs can achieve experimental binding beyond computational evaluation. These additional experiments using other design model require substantial time and resources and cannot be completed within the rebuttal period.

Q2: Reliability of Reduced Diffusion Steps

To assess whether reduced diffusion depth compromises structural accuracy, we evaluated Boltz-2 with T = 25 and T = 200 on the FoldBench antibody–antigen benchmark [1]. Consistent with our design setting, we provided the experimental antigen structure as a template, without an antibody template or MSAs for either partner.

MetricT = 25T = 200
Success rate (DockQ ≥ 0.23) ↑42.07%43.29%
Mean DockQ ↑0.31390.3297
Mean interface RMSD (Å) ↓8.0877.983
Mean ligand RMSD (Å) ↓26.96725.823

T = 25 achieves performance comparable to T = 200, with only a 1.22-percentage-point reduction in Top-1 success rate and modest differences in structural accuracy. Meanwhile, T = 200 incurs 1.72× higher runtime.

These results suggest that T = 25 provides a practical trade-off between prediction accuracy and computational efficiency in this setting, supporting its use during iterative optimization. We included this comparison in the revised Appendix G.2.

Q3: Relationship to Gradient Checkpointing

While both gradient checkpointing and our discrete adjoint formulation reduce memory consumption through recomputation, our approach provides explicit control over differentiation through the diffusion trajectory, with several practical advantages: Decoupled forward and backward computation, Fine-grained gradient control. Please refer to Q2 of our global rebuttal for our detailed response.

[1] Xu, Sheng, et al. "Benchmarking all-atom biomolecular structure prediction with FoldBench." Nature Communications 17.1 (2025): 442.

Q4: Comparison Under Baseline-Specific Evaluation Criteria

Existing antibody design methods generally employ method-specific filtering pipelines before experimental validation, rather than reporting standardized computational success rates across diverse targets. This makes their reported outcomes difficult to compare directly and motivates our 12-target benchmark.

To further assess whether our improvements depend on the evaluation protocol, we additionally evaluated Germinal [2] and AbPilot using Germinal's original 15 filtering criteria on PD-L1, the only target shared between its reported evaluation and our benchmark.

Under the same 24-hour generation budget, Germinal's official implementation with its default configuration generates 95 candidates, none of which pass all 15 filters (https://github.com/SantiagoMille/germinal/blob/main/configs/filter/final/vhh_pdl1.yaml). In contrast, AbPilot achieves the following results:

Filtering criterionGerminalAbPilot + AF3 rect.AbPilotAbPilot + PX rect.
Candidates (n)95453702527
ipSAE1.1%6.8%3.8%7.6%
ipTM4.2%15.0%8.3%15.0%
pTM6.3%23.2%21.7%24.7%
pDockQ222.1%43.3%50.4%40.2%
surfHyd60.0%45.5%45.6%49.0%
PAE25.3%61.1%64.7%53.9%
scRMSD42.1%76.2%71.2%66.4%
shapeC67.4%81.2%77.4%75.5%
hbonds75.8%83.0%68.5%85.8%
ifHyd71.6%69.8%91.5%85.8%
cdr3HS87.4%98.5%99.6%93.5%
nearHS85.3%99.3%99.9%100.0%
clash100.0%100.0%100.0%100.0%
%cdr84.2%100.0%100.0%100.0%
pLDDT77.9%100.0%100.0%100.0%
All 15 filters passed0.00% (0/95)0.44% (2/453)0.14% (1/702)3.23% (17/527)

Each entry reports the percentage of generated candidates satisfying the corresponding Germinal filtering criterion. The final row reports the proportion satisfying all 15 criteria simultaneously. AbPilot also achieves higher pass rates on most individual filtering criteria.

These results show that AbPilot's advantage is not confined to our proposed evaluation protocol: it also produces candidates satisfying the baseline's original filtering requirements under a matched generation budget.

Although this comparison is limited to one shared target, it provides complementary evidence that the observed improvements are not solely attributable to our evaluation criteria. New results are provided in Q2 of the new response to the reviewer CYuq

[2] Mille-Fragoso, Luis S., et al. "Efficient generation of epitope-targeted antibodies with Germinal." Nature biotechnology (2026): 1-10.

Dear Reviewer,

We would appreciate it if you could take a look at our rebuttal and let us know whether your concerns have been adequately addressed. As the discussion period is approaching its end, we would particularly appreciate any further feedback you may have.

We would be happy to clarify any remaining questions.

Best regards,
The Authors

Reviewer_CYuqOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents AbPilot, a framework for antibody design through a frozen diffusion-based co-folding model. It optimizes CDR sequence logits using confidence objectives and geometric objectives defined on the sampled complex. Its main technical contribution is a discrete-adjoint implementation that replays denoising steps and their recorded randomness during backpropagation, avoiding the need to retain activations across the entire diffusion trajectory. Periodic discretization and external-predictor rectification are used to refine the search. Experiments cover 12 targets and both VHH and scFv formats. Prospective validation against mICAM-1 identifies six ELISA-positive candidates among 50 tested designs, with the strongest candidate exhibiting a BLI-measured affinity of 65.8 nM.

Critical strengths
  1. The method addresses a practical limitation of gradient-based design with modern co-folding models. Differentiating geometric objectives through the sampled coordinates provides more direct control over the generated interface than optimizing a trunk-level distance surrogate. The replay-based implementation makes this feasible without retaining the full denoising computation graph. Although the underlying adjoint principle is established, its implementation for a stochastic all-atom co-folding sampler is a useful contribution. The distinction between constant denoiser activation memory and the remaining trajectory storage is also appropriately stated.

  2. The empirical evidence supports the usefulness of the complete approach. Beyond the large improvements over external baselines, the controlled comparison in Table 5 is particularly informative: diffusion differentiation achieves a 30.7% computational success rate, compared with 19.9% for the distogram surrogate and 20.7% for the shortcut variant. This more directly supports the proposed gradient pathway. The prospective identification of an experimentally binding VHH further demonstrates that the approach can produce useful candidates beyond improving computational scores.

Critical weaknesses
  1. AF3 is used for both candidate selection and final evaluation, making it difficult to separate reduced predictor hacking from adaptation to AF3. The Protenix results partially address this concern, while the single-target wet-lab study does not isolate the benefit of rectification.

  2. The relaxed-to-discrete optimization gap remains underexplored. Gradients are computed on soft sequences with a fixed atom layout, so improvements may not persist after discretization. The gradient correctness check on a small surrogate also leaves the actual Boltz-2 implementation less directly validated.

Questions for authors
  1. Given AF3’s role in both rectification and final evaluation, could the authors clarify how the existing cross-predictor results support reduced predictor hacking, beyond improved scores under the selection model?

  2. Do improvements in the relaxed objective predict better discretized candidates when re-evaluated with fresh diffusion seeds? Any existing analysis of this relationship would help clarify the optimization mechanism.

  3. Please clarify which screening, rectification, and final evaluation costs are included in the 24 GPU-hour budget, and report the typical number of completed outer-loop rounds and the main runtime breakdown.

8 comments

We thank the reviewer for their thorough and constructive feedback. We address each point below.

W1 & Q1: Cross-Predictor Robustness

Our original submission already includes cross-predictor validation, where Protenix V1 is used for rectification and AF3 for final evaluation. This improves VHH success from 21.6% to 26.1%, indicating that the benefits of rectification extend beyond the predictor used for candidate selection.

To further assess predictor-specific biases, we additionally evaluated all methods using Mol-1 [1], an independent co-folding model, across the same 12 targets and under identical success criteria. AbPilot achieves 27.3% success for VHH and 24.1% for scFv, substantially outperforming the strongest baselines (2.4% and 1.9%, respectively).

Importantly, under Mol-1 evaluation, VHH success increases from 21.5% without rectification to 24.1% with Protenix V1 rectification and 27.3% with AF3 rectification. These results demonstrate that the improvements generalize across different evaluation models rather than being confined to the predictor used during rectification.

These findings provide stronger evidence for the robustness of AbPilot and reduce concerns about predictor-specific overfitting. We included the complete Mol-1 results in the revised Appendix K.

MethodVHHscFv
AbPilot + AF3 rect.27.3%24.1%
AbPilot (no rect.)21.5%23.9%
AbPilot + PX rect.24.1%20.8%
BoltzGen2.4%1.4%
Germinal1.4%1.9%
RFantibody0.6%0.5%

W2 & Q2: Relaxed-to-Discrete Optimization Gap

We have added a dedicated analysis of the relaxed-to-discrete optimization gap in the revised Appendix E , examining optimization trajectories and discrepancies between soft-sequence predictions and discretized candidates. Please refer to Q1 of our global response for our detailed response.

Q3: Computational Budget and Runtime Breakdown

The reported 24 GPU-hour budget covers the complete design process for each benchmark case on a single NVIDIA L40S GPU, including candidate generation, Boltz-based screening, and in-loop AF3 rectification. Final AF3 evaluation with MSAs is performed separately and is not included in this budget.

Based on runtime logs from the 12 benchmark cases, AbPilot completes an average of 155.1 outer-loop rounds per 24-hour design run. The average computational cost is distributed as follows, we included the results in revised Appendix K:

StageGPU-hoursShare
Candidate generation12.1850.8%
Boltz-based screening and filtering2.5910.8%
In-loop AF3 rectification9.1037.9%
Initialization and other overhead0.120.5%
Total design budget24.00100%
Final AF3 evaluation with MSAs (excluded)8.60—
Total including final evaluation32.60—

[1] Mol-1: https://github.com/molence/Mol-1

[2] Boyd, N., Guns, S. & Escalante Bio. Mosaic. https://github.com/escalante-bio/mosaic (2025).

Reviewer_CYuqOpenReview ↗

Thanks for the Response

Thanks for the detailed response. Q3 is resolved, Q2 is largely addressed by Appendix E, and the Mol-1 evaluation is the right experiment for W1. The following now dominate my assessment.

1. The headline success rates are conditional on screening. Per your response to fJny, the denominator is the Boltz-2-filtered count and the numerator the AF3-passing count, and BoltzGen is additionally cut to its top 1,000 per target. This measures success among retained candidates and does not by itself show the overall yield advantage under the matched budget. Please report, per method, the numbers generated, passing the screen, evaluated by AF3 and passing the final criteria, then final passes per generated candidate and per design GPU-hour (final AF3 cost reported separately). These counts should be recoverable from existing logs. I weight this most heavily, since the paper's central claim is a magnitude.

2. The PD-L1 comparison does not by itself establish a pass-rate advantage. The all-filter counts are 0/95 (Germinal), 1/702 (base), 2/453 (AF3 rect.) and 17/527 (Protenix rect.). Germinal's 0/95 has an exact 95% interval of about [0, 3.8%], which contains all three AbPilot rates; only the Protenix variant is borderline against it (Fisher one-sided p ≈ 0.06), and it is not the configuration behind your headline numbers. Please report confidence intervals. Separately, this matched-budget run does not show whether Germinal's published performance is reproduced under its original protocol.

3. The per-criterion results show gains alongside a trade-off. I recognize the improvements in scRMSD and shape complementarity, and the geometric analysis in your reply to fJny. However, surface hydrophobicity is worse for all three variants (45.5% to 49.0% versus 60.0%), and hbonds and ifHyd are mixed. Please discuss this trade-off, including developability, and what the confidence-derived and geometric results establish about reduced predictor-specific bias.

4. Mosaic (Matched) needs diagnostics or a narrower interpretation. A result of 0.0% in both formats, versus 4.1% and 4.7% for Mosaic (Default), does not identify gradient conflict as the cause, since transferring your loss weights while keeping the other Mosaic settings may simply be sensitive to those settings. Please provide loss curves and per-target counts, or label this as an untuned objective-transfer ablation.

5. The contribution framing should reflect the checkpointing clarification. Your reply to YkJC states that the method does not compute a fundamentally different gradient from exact checkpointed backpropagation. Please carry this into the abstract and contribution list, relating the O(1) activation memory to stepwise recomputation and foregrounding application, integration and gradient control. Please also state which AlignProp configuration the 7.1% figure is measured against.

Item 1 is the one I would most like answered.

AuthorsOpenReview ↗

Response (1/3)

Q1:

Thank you for raising this point. We believe this concern stems from a misunderstanding of how candidates are defined in our design pipeline. Boltz-2 screening is an integral part of AbPilot’s design procedure, not an additional post hoc filter applied to finalized outputs. Candidates rejected during this procedure are intermediate search attempts; the retained candidates constitute the algorithm’s final outputs. All internal generation, prediction, and screening are included in the same 24-hour design budget used for the comparison.

This output convention is not specific to AbPilot. Other generative models (e.g. BindCraft, Germinal) likewise define their accepted outputs as designs that pass their internal computational filters. Therefore all the models are compared under the same budget covering all actions in their own systems, including generation and internal filters.

Accordingly, the denominator in our reported success rate is the number of final candidates emitted by the design pipeline, and the numerator is the number satisfying the final AF3 criteria. AbPilot produces a substantial number of such candidates within the allocated budget. The reported rate therefore measures the quality of the algorithm’s final outputs, rather than all intermediate search attempts.

To conclude, our comparison thus evaluates final pipeline outputs under the same design-time constraint, without granting AbPilot additional computation for screening. Each method receives one GPU for 24 hours per target, corresponding to 288 design GPU-hours across the 12 targets. Final successful designs per GPU-hour constitute a complementary measure of productivity, rather than a correction to this output-level success-rate definition. Every candidate counted in #output underwent AF3 evaluation; therefore, #AF3 evaluated equals #output for every method. We clarified the candidate definitions, the role of internal screening, and its inclusion in the 24-hour budget to prevent this ambiguity in our paper.

For further confirmation, we report the following metrics:

  • #gen: raw number of generations. This is not countable for Germinal as the implementation only retains those passing the internal filters, so the intermediate candidates are not saved to disks.
  • #output: raw number of outputs by each generative system, namely candidates passing the internal filters of each system
  • #succ: successful candidates by AF3 evaluation
  • pass/gen%: successful candidates per generation in percentage
  • pass/output%: successful candidates per output in percentage
  • pass/GPU-h: successful candidates per gpu hour
  • design GPU-h: total GPU hours used by the generation & internal filter phase, which are 24h for each of the 12 targets
  • AF3 GPU-h: total GPU hours used by later AF3 evaluation
FormatMethod#gen#output#succpass/gen %pass/output %pass/GPU-hdesign GPU-hAF3 GPU-h
VHHBoltzGen39,35412,0001500.381.250.52288206
VHHGerminal--1164--3.450.012885
VHHRFantibody5,1344,974220.430.440.0828889
VHHAbPilot14,4123,9909446.5523.663.2828872
VHHAbPilot + AF3 rect.8,1722,74292611.3333.773.2228852
VHHAbPilot + PX rect.9,7414,2321,13811.6826.893.9528876
scFvBoltzGen32,09212,0001690.531.410.59288242
scFvGerminal--38427--7.030.092888
scFvRFantibody9,9498,546410.410.480.14288173
scFvAbPilot6,8303,34292513.5427.683.2128868
scFvAbPilot + AF3 rect.3,9612,20263516.0328.842.2028845
scFvAbPilot + PX rect.5,1212,90482616.1328.442.8728859

Comments on the paper

2 comments
AuthorsOpenReview ↗

Global response

Q1: Relaxed-to-Discrete Optimization Gap

We have added a dedicated analysis of the relaxed-to-discrete optimization gap , examining both optimization trajectories and the discrepancy between soft-sequence scores and discretized candidates in revised Appendix E.

Optimization procedure. In the first round, we initialize the sequence by uniform random sampling and construct its corresponding atom layout. Subsequent rounds are initialized from the best discrete candidate selected in the preceding round. Within each round, the atom layout remains fixed while sequence probabilities are optimized. The layout is updated only between rounds, based on the selected discrete sequence.

Optimization trajectories. We analyzed the Boltz metrics recorded during optimization across 12 targets (Figure 5 a-c). Trajectories were aligned by relative optimization progress, with the median of target-level means reported. The results show consistent overall improvements in key optimization metrics. These findings confirm that the relaxed objectives can be effectively optimized using the actual Boltz implementation.

To assess whether relaxed predictions remain informative after discretization, we paired the soft-sequence metrics at the minimum total loss checkpoint of each round with their discrete candidate from the same round. We observe strong positive rank correlations (Figure 5 d-f). These results indicate that relaxed-state metrics preserve substantial ranking information about discrete designs across optimization rounds.

Soft-to-discrete discrepancy. We further investigated how the number of inner-loop optimization steps affects the gap between relaxed and discrete evaluations (Figure 5 g-i). For each round, we compared the soft-sequence metrics at the minimum-total-loss checkpoint with the average metrics of five corresponding discretized candidates after Boltz re-prediction.

We observe a systematic optimism gap. This gap generally increases from 10 to 100 inner-loop steps, motivating our choice of 25 steps as a practical balance between optimization progress and discretization error.

For additional context, we evaluated Mosaic [1] in the same setting, which performs optimization using sequences containing the placeholder token X. Mosaic exhibits a larger soft-to-discrete gap under the evaluated settings.

Overall, these results highlight the importance of explicitly accounting for the relaxed-to-discrete gap. Rather than relying solely on optimized soft-sequence scores, AbPilot mitigates their potential bias through periodic discretization, independent structure re-prediction, and selection of the best discrete candidate to initialize subsequent optimization rounds.

Q2: Relationship to Gradient Checkpointing

While both gradient checkpointing and our discrete adjoint formulation reduce memory consumption through recomputation, our approach provides explicit control over differentiation through the diffusion trajectory, with several practical advantages.

Decoupled forward and backward computation. Rather than retaining a computational graph across denoising steps, we separate forward sampling from backward differentiation and explicitly replay individual transitions to compute local vector–Jacobian products. This provides direct control over the intermediate states and computations used during backpropagation.

Fine-grained gradient control. By exposing gradient contributions from individual denoising steps and separating diffusion and confidence pathways, our formulation also enables more flexible analysis and manipulation of gradient propagation, including step-dependent weighting and gradient conflict mitigation (Figure 7f).

Crucially, our contribution extends beyond memory-efficient differentiation. We integrate trajectory-level differentiation into an end-to-end antibody optimization framework, enabling direct optimization of structure-dependent geometric objectives through diffusion co-folding models, rather than relying exclusively on shortcut differentiation or distogram-based surrogates.

We clarified the relationship to standard gradient checkpointing in the revised Appendix B.

[1] Boyd, N., Guns, S. & Escalante Bio. Mosaic. https://github.com/escalante-bio/mosaic (2025).

AuthorsOpenReview ↗

Summary of Revisions for the Area Chair and Reviewers

Dear Area Chair and Reviewers,

Thank you for your constructive feedback. We have carefully addressed the main concerns and conducted extensive additional experiments to strengthen our evaluation and clarify the contributions of AbPilot. Our new results demonstrate that:

  • AbPilot substantially outperforms Mosaic on the 12-target benchmark, achieving success rates of 21.6% vs. 4.1% for VHH and 26.0% vs. 4.7% for scFv, respectively.
  • Full-trajectory differentiation preserves high gradient fidelity, achieving a median relative error of 6.98 × 10⁻⁷, while reducing peak GPU memory by 7.1% compared with our Gaussian-truncated AlignProp adaptation.
  • AbPilot enables effective epitope-specific geometric control, with 73.85% of evaluated designs satisfying both epitope attraction and framework avoidance constraints after AF3 re-prediction, compared with 21.73% and 45.51% for alternative differentiation strategies.
  • Independent cross-predictor evaluation confirms robust performance, with AbPilot + AF3 rectification achieving 27.3% VHH and 24.1% scFv success under Mol-1 evaluation, substantially outperforming existing baselines.
  • Additional analyses validate the optimization framework, including relaxed-to-discrete consistency, diffusion-step sensitivity, matched-budget comparisons, and detailed computational cost breakdowns.

We have also expanded the implementation details, clarified the evaluation protocols, and revised the manuscript to better distinguish our methodological contributions from existing differentiable diffusion approaches.

We hope these additional results address your concerns and strengthen the evidence supporting our contributions. We would greatly appreciate your consideration of our rebuttal and welcome any further feedback.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.