acceptodds
Under review as a conference paper at ICLR 2027

RAFA: Reward-Aligned Flow Tuning for Antibody and Nanobody Design

13 views

Abstract

We introduce RAFA, an antigen-conditioned framework that jointly generates CDR sequences and the complete antibody structure from an antigen structure, epitope information, and an antibody framework sequence, without requiring initial antibody coordinates. Antibody and nanobody design requires coordinating sequence, structure, and the interface with a specified antigen. Generative models trained to reconstruct observed complexes do not necessarily optimize the quality of their generated interfaces. Building on La-Proteina's partially latent autoencoder, the model combines explicit coordinates with per-residue latents encoding sequence and side-chain geometry. Joint flow matching is complemented by geometry supervision. We further introduce Reward-Gradient Transport (RGT), which approximately transports terminal reward gradients with respect to antibody coordinates to recorded states along the same sampled trajectory and updates the shared generative velocity field. Relative to RAFA, RAFA (RL) improves mean DockQ by 22.0% for antibody design and 28.9% for nanobody design on a globally filtered, held-out benchmark. Additional structure-prediction, refolding-confidence, and interface analyses characterize both improvements and trade-offs. These results support coordinate reward-gradient post-training as an approach to improving computational antibody–antigen interface generation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
antibody design, flow matching, reward alignment, sequence–structure co-design, protein complex generation

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 3 reviews, above 85% of the others.
Reviewer_pDpTOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes RAFA, a framework for the joint generation of antibody and nanobody sequences and structures conditioned on a given antigen. It employs Reward-Gradient Transport (RGT) to incorporate terminal coordinate-based reward gradients into the update of velocity fields—evaluated over sparsely sampled states—thereby improving the quality of the generated interfaces at a lower computational cost.

Critical strengths

The experiments cover structure prediction and design for two types of antibody formats, incorporating comprehensive mechanistic controls. Post-training RGT on VH–VL pairs yields performance gains that extend to VHH formats; furthermore, sequence refolding using Protenix-v2—a model distinct from the source of the reward labels—results in continued improvements in DockQ scores. Under identical experimental conditions, both RGT and full BPTT achieve an average DockQ score of 0.172, yet RGT reduces training time from 55 hours to 24 hours and peak VRAM usage from 72 GiB to 24 GiB (Table 9), demonstrating its practical efficiency.

Critical weaknesses
  1. Empirical support for the incremental benefits of the proposed method remains insufficient. While Appendix F.4 discusses the shared structure and differences between RGT and DRTune, the closest "matched" DRTune baseline is presented only graphically in Figure 4, lacking precise numerical values ​​and an uncertainty analysis of the paired differences. In Table 9A, the DockQ gains for linear time-weighting—relative to constant and quadratic weighting—are 0.008 and 0.003, respectively; furthermore, the inference-seed repetitions in Table 10 do not sufficiently demonstrate the robustness of these differences against the stochasticity inherent in the training process. Given that the "random" and "shuffled" controls in Table 8 also show some improvement, it is necessary to more accurately quantify the specific gains attributable to correct gradient directions and trajectory pairing. It is recommended that the authors provide paired confidence intervals for key baselines, conduct additional independent training runs, and clarify the number of training seeds used for the results in Table 9B.

  2. The benefits relative to filtering at inference time remain unclear. The paper lacks a direct comparison against a "best-of-N" re-ranking approach, where the same reward model is used to select from outputs generated by the base model. It is recommended to report the quality—as well as the generation and scoring costs—for both RAFA and RGT using identical candidate budgets (e.g., N=1N=1 and N=8N=8). This would clarify how the post-training process improves the quality-cost trade-off. Even if a "best-of-8" selection from the base model performs comparably to a single-sample output from RGT, the significance of this result should be evaluated by accounting for the additional inference costs alongside the amortized training costs.

Questions for authors
  1. The reward relies primarily on interface geometric proxies and a specified epitope; therefore, its correlation with docking metrics should be clearly qualified in the conclusions. Independent refolding results serve as valuable supplementary evidence; please clarify whether the MSA, templates, random seeds, and candidate selection protocols used for refolding were consistent across all methods. It is recommended to report the direct sequence identity between the sequences generated by RAFA and RGT to help explain how sequence variations contribute to improvements in refolding; similar amino acid recovery (AAR) values ​​alone are insufficient to address this.

2 Please provide the predictive performance of the fold-confidence surrogate on held-out calibration samples to demonstrate the validity of this reward component. Regarding the lower DockQ scores for DiffAb/MEAN in Table 14, please also provide information on the quality of the initial complexes or include an appropriate sanity check to distinguish the impact of the initial poses from that of the subsequent design steps.

2 comments

Dear Reviewer,

Thank you again for your valuable comments. As the discussion period is nearing its end, we would like to kindly ask whether our responses have adequately addressed your concerns. Please let us know if any further clarification is needed. If our responses have resolved your concerns, we would sincerely appreciate your consideration in updating your evaluations accordingly.

Dear Reviewer pDpT,

We would appreciate a brief update on whether our additional experiments and clarifications address your concerns, particularly the matched-budget comparisons, best-of-N evaluation, and surrogate validation. If any concerns remain, we would welcome the opportunity to clarify them before the discussion closes.

Thank you again for your time and consideration.

Best regards, The Authors

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

This paper presents RAFA, a flow-based framework for antigen-conditioned antibody sequence-structure co-design. RAFA is built on the La-Proteina framework, with antigen structure representation injected as a condition. The authors also introduce reward-gradient transport, a post-training/RL mechanism that transports the interface rewards of the terminal state back to the intermediate state of the generative process and reshape the local velocity. Benchmarking on 93 antibody test set against established methods exhibits comparable structural prediction performance and superior geometric, energetic and docking metrics in de novo generation, especially with RL. The authors also show generalizability to nanobody designs. Ablation study demonstrates the importance of both the structure/docking-based reward and the general protein structure data, as well as the effectiveness of the RGT post-training.

Critical strengths
  1. The study moves beyond traditional evaluations such as sequence/structure recovery or model-predicted foldability and positions structural and biophysical metrics as a critical guidance signal for the generation quality. The RGT velocity alignment effectively translates the multi-modal rewards into the generative process. This could be highly informative for the field as a step forward bridging the gap between data-driven probabilistic modeling and universal biophysical rules often not directly learnable from data.
  2. The ablation study is particularly revealing. It highlights how the incorporation of general protein structure knowledge and external structural/biophysical knowledge can produce more physically favored generations. Thus, the proposed framework provides an effective means of integrating prior knowledge about general protein folding and interaction.
Critical weaknesses

Since the base RAFA actively optimizes an auxiliary reward that contains backbone geometry, interface geometry and contact recovery, etc. during training, the overall improvement in the evaluation metrics is somewhat expected. Other methods may benefit from a similar setup. This may benefit from some further clarification, see Questions.

Questions for authors
  • What is the main advantage of the teacher-student training compared to simpler conditioned generator with the antigen structure provided?
  • Standardization of terminology is needed: what are the "structural loss" and "docking loss" ablated in the base training in Table 6? Which parts do they correspond to the loss components in Appendix C5?
  • C$\alpha$ geometry seems rather coarse for the analysis of contact and interface geometry. Please briefly discuss how much those C$\alpha$-based interface loss components add to the model performance.
  • How does ablating certain parts of the RGT reward in Eq. (6) affect the performance?
Reviewer_57MhOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RAFA for antigen-conditioned antibody sequence and structure generation. RAFA builds on La-Proteina with additional geometry supervision. The authors also propose Reward-Gradient Transport (RGT), which uses terminal reward gradients to update intermediate flow velocities without backpropagating through the sampling trajectory. Experiments show that RGT improves DockQ by 22.0% for antibodies and 28.9% for nanobodies.

Critical strengths

The proposed post-training algorithm is technically sound, and its BPTT-like formulation appears computationally efficient.

Critical weaknesses

The post-training algorithm lacks a direct comparison with other differentiable optimization methods, such as Adjoint Matching [1].

[1] Carles Domingo-Enrich, Michal Drozdzal, Brian Karrer, and Ricky T. Q. Chen. Adjoint Matching: Fine-tuning Flow and Diffusion Generative Models with Memoryless Stochastic Optimal Control. International Conference on Learning Representations (ICLR), 2025.

Questions for authors
  1. Can the authors compare RGT with Adjoint Matching [1] under the same computational budget?

  2. How accurate is the identity approximation used in RGT? Can the authors empirically compare its gradients with exact BPTT?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.