acceptodds
Under review as a conference paper at ICLR 2027

PODA: Privileged On-Policy Distillation for Antibody and Nanobody Design

Abstract

Epitope-conditioned de novo antibody design seeks to generate CDR sequences and the corresponding antibody–antigen complex structure for a specified target site. Flow-matching models learn from interpolations toward native complexes, but generate through their own predicted states, creating a potential mismatch between training and inference. We introduce PODA (Privileged On-Policy Distillation for Antibody and Nanobody Design), a two-stage framework for joint CDR sequence and complex-structure generation. Given an antigen sequence and structure, a specified epitope, and a framework sequence with masked CDRs, PODA first trains a deployable student and a privileged teacher with supervised objectives on the same complexes. It then freezes the teacher and updates the student on states obtained by re-noising endpoints generated online by the evolving student policy. Distillation matches teacher-predicted clean endpoints, with residue-level weighting emphasizing CDR and interface geometry and a quality gate selecting teacher targets according to their reconstruction error. On the sequence-filtered test set, PODA increases the design success rate from 20.4% to 32.3% on antibody design and from 12.0% to 24.0% on nanobody design relative to the supervised base model. Ablations support the contribution of privileged supervision, and PODA achieves the highest mean DockQ among the evaluated post-training methods in both cohorts. These findings support privileged endpoint supervision on student-induced states as an effective approach to improving complex interface geometry.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Antibody Design, Sequence–Structure Co-design, Flow Matching, On-Policy Distillation, Privileged Information

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_DAvQOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents PODA (Privileged On-Policy Distillation for Antibody and Nanobody Design). This is a two-stage antibody sequence–structure flow framework. In Stage I, a public-input student and a privileged teacher are jointly supervised on the same complexes, yielding the resulting student model. In Stage II, the teacher is frozen while the student is updated on re-noised endpoints generated online by its evolving policy. The distillation process aligns the student with teacher-predicted clean endpoints, places emphasis on CDR and interface residues, and gates teacher targets according to reconstruction quality.

Critical strengths
  1. The paper proposes a distillation method based on a flow network model for conditional antibody CDR design. Through a two-stage training process, the method successfully achieves the goal of antibody CDR generation.

  2. Antibody design is a highly complex challenge in practical drug development. In this paper's experimental evaluation, the authors incorporate a broad range of validation metrics to assess performance as comprehensively as possible, which is a commendable approach.

Critical weaknesses
  1. In terms of algorithmic innovation, the reconstruction objective used to train the flow model in the first stage is already widely adopted, while the distillation of information between two models with identical architectures in the second stage is also a well-established approach. Consequently, the paper offers limited technical novelty overall.

  2. Regarding experimental evaluation, antibody CDR design is a well-established research topic with numerous relevant baseline methods based on GNNs, diffusion models, and flow-based models. The paper’s limited incorporation of baselines hinders a rigorous assessment of the proposed method’s advantages.

  3. Importantly, the ultimate goal of AI-assisted antibody design is to advance candidates into real-world clinical drug development pipelines. Although this study does not include wet-lab experiments, it should provide sufficiently robust in silico validation to inform future experimental work. However, the scope of the evaluation remains limited: CDR segments typically contain only around 10 amino acids, and the test set comprises fewer than 100 antibodies. These limitations substantially constrain this study’s research value and potential impact.

Questions for authors

The authors should more clearly articulate this paper’s principal algorithmic contributions and expand the experimental evaluation to address its current limitations, thereby providing more meaningful insights for future research and practical applications.

Reviewer_E1SdOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PODA (Privileged On-Policy Distillation for Antibody and Nanobody Design), a two-stage sequence–structure flow framework for epitope-conditioned antibody and nanobody design. The deployment inputs are an antigen structure, a specified epitope mask, and an antibody framework sequence with CDR residues masked; the model must generate the missing CDR identities and the complete antibody structure in the antigen frame, without antibody coordinates at inference.

The method uses a public-input student and a privileged teacher. In Stage I, both are trained with supervised flow matching on the same complexes; the teacher additionally receives complete antibody sequence, native antibody coordinates, and interface annotations. The resulting student is PODA (base). In Stage II, the teacher is frozen, and the student is updated on states obtained by re-noising endpoints generated online by its own evolving policy. The teacher supplies clean-endpoint targets, with CDR/interface-aware weighting and a teacher-quality gate; an auxiliary RAM correction and a supervised anchor are also used. At inference, only the student and public inputs are required.

The main contributions are: (1) a unified design pipeline for VH–VL antibodies and VHH nanobodies under the same framework-sequence-conditioned formulation; (2) an on-policy privileged distillation scheme that transfers training-only structural supervision without changing deployment inputs or search budget; and (3) a systematic empirical validation on 93 VH–VL and 50 VHH test targets, including design comparisons, post-training comparisons, training-stage and teacher ablations, alternative post-training methods, and complementary structural diagnostics.

Critical strengths

The paper evaluates all 93 VH–VL and 50 VHH test targets, uses a temporal split with sequence-based redundancy filtering, reports multiple structural and interface metrics, includes five-seed design results, and compares against several post-training methods. The side-chain/interface analysis (Table 3) and complex-prediction setting (Table 5) go beyond the usual DockQ-only reporting.

Critical weaknesses

The improvement is real but narrow, and the paper sometimes overstates it relative to dedicated structure predictors. PODA improves DockQ and SR over PODA (base), but global reconstruction and fold metrics often remain essentially unchanged or slightly worse (e.g., Table 4: AAR-H3 0.512 vs 0.509, TM-score 0.927 vs 0.925, IDDT 0.780 vs 0.778 for PODA (w/oRAM) versus PODA (base)). In the complex-prediction setting, PODA does not surpass Proteinx-v2 or AlphaFold 3 in mean DockQ on VH–VL or VHH targets, although it has higher SR in some comparisons. The paper acknowledges different rankings, but the main claim should be scoped more carefully: the evidence supports improved interface placement under the design setting, not uniformly better structural quality.

Questions for authors
  1. Can you provide a comparison under matched input contracts—e.g., PODA with predicted framework coordinates versus RFantibody/DiffAb with initialized framework structures, and PODA with framework sequence only versus sequence-conditioned baselines?

  2. Table 15 stratifies by antigen identity, but not by antibody sequence or CDR-family identity. Can you report performance stratified by antibody/CDR similarity to training data, and ideally under a strict antibody-family holdout?

  3. How different are the teacher’s clean-endpoint predictions from the native endpoints, and how often does the quality gate reject teacher targets? Since Direct GT is weaker than teacher-guided distillation, it would be informative to quantify the teacher’s endpoint error relative to the student and native endpoints, and to analyze whether the gate primarily selects easy examples or genuinely corrective teacher predictions.

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

This paper studies epitope-conditioned de novo antibody and nanobody design: given an antigen sequence and structure, a specified epitope, and a framework sequence with masked CDRs, the model must generate CDR identities together with the complete antibody structure in the antigen frame, without access to antibody coordinates or the native pose. The authors target the mismatch between flow-matching training, which samples states along paths toward ground-truth complexes, and inference, which recursively follows the model's own predictions. PODA proceeds in two stages. First, a deployable student and a privileged teacher, the latter conditioned on additional training-only information, are trained with supervised objectives on the same complexes. Second, the teacher is frozen and the student is updated on states obtained by re-noising endpoints that the evolving student itself generated, matching the teacher's predicted clean endpoints with residue-level weighting toward CDR and interface regions and a quality gate that rejects unreliable teacher targets. An auxiliary reward-regression correction and a supervised anchor are also included. Evaluation covers separate conventional-antibody and nanobody cohorts from a standard structural benchmark, with sequence-recovery, structural, and docking metrics, plus comparisons against several alternative post-training methods.

Critical strengths

The information boundary is defined cleanly and respected throughout. Privileged information enters only the teacher's conditioning and the loss weighting; the student's deployment inputs never change, and no teacher or native structure is needed at inference. The paper is careful to restate this in several places.

One finding is interesting and somewhat counterintuitive: distilling a privileged teacher's predictions on student-induced states outperforms regressing the ground-truth endpoints directly on those same states, and a non-privileged teacher is clearly weaker than the privileged one. This pair of controls is a nice piece of experimental design.

The ablation coverage is the most thorough of the three papers I reviewed. It separates the two stages, the presence of the teacher, the supervision target, the state source, and each gating term, and the state-source comparison shows a monotone ordering that supports the on-policy claim. The authors also report that removing the teacher while keeping rollout and re-noising performs worse than the supervised base, which usefully shows that on-policy states alone are not the source of the gain.

The writing is appropriately restrained. The authors repeatedly note that local loop accuracy and complex-level agreement behave differently, state explicitly that the qualitative examples were selected post hoc and are not representative, clarify that the auxiliary score is not a measured binding affinity, and caution that their redundancy filter should not be read as a strict low-homology benchmark. The ethics discussion of dual-use risk is appropriate.

Critical weaknesses

The statistical strength behind the headline result is limited, and the abstract's framing amplifies this. The success-rate improvements correspond to a modest number of additional targets on cohorts of fewer than a hundred and around fifty targets respectively, and the reported multi-seed variation appears to come from inference seeds rather than retrained models. Q1: Could paired per-target tests or bootstrap intervals be reported, and could the post-training stage be repeated across training seeds to show how stable the gain is?

The mean docking score and the success rate move by very different relative amounts across the gating ablations, which suggests a threshold effect. Q2: Is the jump driven by a small number of targets crossing the success threshold? A cumulative distribution of the docking score, or success rates at several thresholds corresponding to different quality tiers, would clarify whether the whole distribution shifts.

Absolute performance remains low, with most designs falling short of what would usually be considered an acceptable interface. Q3: Could the fraction of medium and high quality predictions be reported alongside the current success criterion? I would also note that the Rosetta-style interface energy improvements are weak evidence of function, which the ethics statement already acknowledges and which could be stated equally clearly in the results discussion.

Compute budgets are not aligned between the compared settings, since the post-training stage requires online rollouts while evaluation gives every method a single candidate per target. Q4: If the supervised base model were given a comparable inference-time budget, for example sampling several candidates and selecting with a deployable scorer, how would it compare? The paper cites this inference-time scaling route in related work but does not test it.

Test-set redundancy control is acknowledged as weak, but its consequences are not fully quantified; a large share of targets sit in the higher-similarity stratum, and in the lowest-similarity stratum the metrics move in inconsistent directions on a small number of targets. Q5: How should the divergence between improved success rate and worsened interface RMSD in that stratum be interpreted, and is the sample size there sufficient to support conclusions about generalization?

Baseline comparability is my largest remaining doubt. Several baselines are systems built around large pretrained weights and their own multi-candidate sampling and filtering pipelines, and some of their reported numbers here look far below what their original papers describe. Q6: Were these baselines retrained under the shared setting or run from official weights, and were pipelines that normally depend on heavy sampling and filtering disadvantaged by the single-candidate protocol? Making each baseline's exact protocol explicit seems necessary for the main tables to be read as a fair comparison.

The attribution of the gain is also not fully clean, since the objective includes an auxiliary correction borrowed from one of the compared post-training methods, and removing it accounts for a visible part of the improvement. Q7: How much does privileged on-policy distillation contribute on its own, and should the abstract attribute the gain to the combination rather than to distillation alone?

Finally, one appendix table appears to show a non-trivial drop in a global structure metric for the gated variant on the nanobody cohort, which is larger than the "modest changes in fold scores" described in the text. Q8: Is this real, and does gating trade global fold quality for interface geometry? Separately, since the most closely related prior work already applies privileged on-policy distillation in a narrower setting, a direct experimental comparison, or an explanation of why one is not feasible, would strengthen the positioning.

Questions for authors

see waekness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.