acceptodds
Under review as a conference paper at ICLR 2027

EpiForge: Risk-Averse Verifier-Guided Editing for Epitope-Conditioned Antibody Design

Abstract

Epitope-conditioned antibody design is dominated by open-loop pipelines that generate candidates, score them once with a structure predictor, and discard whatever fails, so a rejected candidate never informs the next proposal, developability is inspected only after the fact, and a design that pleases one folding oracle need not survive another. We present **EpiForge**, a closed-loop editing agent that reframes the task as verified repair of an initialized variable region. An epitope-conditioned initializer writes CDR sequences onto a fixed framework by masked discrete diffusion over a chemistry-aware coarse-grained graph. A reasoning policy then proposes edits as a short integer genotype; a parameter-free mapper projects that genotype onto a legal mutation set by sequential modular arithmetic, so no proposal can be malformed, out of bounds, duplicated or vacuous; an uncertainty-aware world model predicts the resulting interface metrics with its own variance; and a confidence-bounded selector spends the scarce verification budget where it is most informative. Acceptance is not a learned gate: every retained edit is confirmed by a real structure prediction and the metrics computed from it, and anything that does not improve the verified state is rolled back. Selection is optimistic and credit assignment conservative, since the selector adds the surrogate's predicted standard deviation while the group-relative objective subtracts it. On a SAbDab split that removes redundancy on the antibody, antigen-sequence and antigen-structure sides at once, EpiForge reaches 45.3% CDR-H3 recovery and 55.1% all-CDR recovery, improves CDR-H3 backbone accuracy under reporting predictors never used inside the loop, and attains the best humanness of every method compared without optimizing for it. A held-out case study recovers pose, loop conformation and sequence together at DockQ 0.649 against 0.387 for the strongest baseline, while our ablation and two further cases show these three quantities coming apart; we report where the loop trades sequence recovery for interface quality, and where it does not lead. Our code is at [https://anonymous.4open.science/r/EpiForge-82C6/](https://anonymous.4open.science/r/EpiForge-82C6/).

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Antibody Design, Epitope-Conditioned Design, Closed-Loop Editing Agent, Verifier-Guided Optimization, Uncertainty-Aware Reinforcement Learning

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_8NpdOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Summary: EpiForge reframes epitope-conditioned antibody design from one-shot generation into closed-loop verified repair: propose small mutations, cheaply rank them, verify promising ones with structure prediction, keep only verified improvements, and learn from both accepted and rejected edits.

Contributions: The key innovations are (1) a verifier-controlled loop where learned models cannot directly accept edits, (2) GAUGE, a deterministic decoder guaranteeing legal 1–3-residue mutations, and (3) risk-aware learning: uncertainty encourages exploration when selecting edits to verify but penalizes uncertain edits when assigning policy credit.

Validation: On a redundancy-controlled SAbDab split, EpiForge achieves 45.3% H3 / 55.1% all-CDR recovery, improves backbone accuracy under independent structure predictors, and improves substantially over its own initializer. Case studies and ablations also show an important nuance: sequence recovery, loop geometry, and interface quality can diverge. The evidence is entirely in silico, single-run, and ultimately dependent on imperfect structural verifiers, so experimental binding/developability remains unvalidated.

Critical strengths
  1. Strong closed-loop formulation with explicit safeguards against surrogate exploitation. EpiForge cleanly separates proposal/ranking from acceptance: learned components determine what is worth testing, but only verified structural improvement can advance the design. The asymmetric use of uncertainty—optimistic for verification and conservative for policy credit—is particularly compelling. This is consequential because it addresses a fundamental weakness of open-loop or surrogate-optimized design rather than merely introducing another generator.

  2. Thoughtful evaluation designed to test whether improvements generalize beyond the optimization signal. The redundancy-controlled split and, especially, evaluation with structure predictors disjoint from the in-loop verifier make the reported gains more credible. EpiForge improves substantially over its own initializer, helping isolate the value of the proposed editing loop. This substantially strengthens my confidence that the reported computational gains are not simply verifier overfitting.

Critical weaknesses
  1. The evidence remains entirely in silico and is strongly verifier-dependent. EpiForge uses a single structural verifier inside the optimization loop, so systematic errors in that verifier can directly shape the designs; the independent reporting predictors can reveal disagreement but cannot prevent it. Moreover, no candidates are synthesized or experimentally assayed. This is consequential because the paper cannot establish that its computational improvements translate into actual binding affinity or developability.
  2. The ablations do not cleanly establish the contribution of the individual components. The module ablations are not verification-budget matched; notably, simpler variants can achieve competitive improvements while using substantially fewer verifier calls. This makes it difficult to determine whether the full SKETCH/ECHO/SCOUT machinery is necessary for the reported gains, weakening the empirical support for the proposed system architecture rather than just the overall closed-loop idea.
Questions for authors
  1. Can the authors provide a verification-budget-matched ablation of the full method against SKETCH-free and ECHO-only variants? The current ablations use substantially different numbers of verifier calls, making it difficult to determine whether the full architecture provides gains beyond additional verification/search budget. This could materially affect my assessment of the necessity of the proposed components.

  2. How robust are the main results to stochastic variation and to the choice of in-loop verifier? The current evaluation reports a single run and uses one in-loop structural verifier. Results across multiple seeds and/or an alternative verifier would help establish whether the improvements are robust rather than dependent on a particular search trajectory or verifier bias. This directly affects my confidence in the empirical conclusions.

Reviewer_FC5FOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

EpiForge is an agent that iteratively edits antibody sequences using integer-coded mutations, a cheap quality/uncertainty predictor, and an expensive structure verification. The verification module controls acceptance, and the rejected evaluations can inform later iterations. Predicted uncertainty encourages exploration. The main contribution is this combination of modules.

Critical strengths
  1. The released source code and dataset split
  2. The simple deterministic edit mapper is an interesting design choice
  3. Evaluation distinguishes recovery, loop shape, and antigen placement.
Critical weaknesses
  1. The manuscript is difficult to follow because of extensive LLM-invented jargon, frequent changes in wording for related concepts, and shifts in how terms are used. Consistent terminology and more direct descriptions of the method and experiments would improve clarity.

  2. The verifier uses the known native structure during search. Section 3.1 states that the inputs are antigen, epitope, framework and pairing, but the reward uses DockQ and H3 RMSD. The released scoring implementation compares H3 coordinates and H3-antigen contacts against native complex and uses the resulting scores for edit acceptance. DockQ also requires a native reference. Evaluating final outputs with different predictors does not remove information from native structures already used during search. So, if the task is reconstruction using known antibody coordinates, it should be stated clearly. Otherwise, a search score without native antibody coordinates and corresponding results on existing targets should be presented.

  3. Experimental setup and reporting need clearer descriptions. For example, the starting designs (stage-0 initizalization) are described as masked discrete sequence diffusion on a coarse graph with an ESM-C encoder in Section 3.2/Appendix A.5 and the main figure. But Table 1/Section 4.1 reads "RFantibody + ProteinMPNN". The descriptions do not establish whether these are different stages, alternative configurations or the same procedure described incompletely. Similarly, the text allows all 6 CDRs, but the released code edits only H3. Another one is the discrepancy between mean H3 RMSD values for dyMEAN/EpiForge in tables 1 and 7. One clear protocol mapping the method sections, code and result tables and allows to identify the initialization procedure, editable CDRs, checkpoint, target set, etc. would strengthen the manuscript.

  4. The complexity of the proposed method is insufficiently justified by ablations. The ablations are limited and done at different budgets.

  5. Diversity is unclear. Table 2 reports substantially higher humanness, H3 recovery and fewer liability motifs than reference native antibodies. Reporting diversity and uniqueness could clarify if there's a mode collapse or not.

Questions for authors
  1. What information is available during search? Does EpiForge require the native antibody structure to calculate its reward? If not, how are DockQ and H3 RMSD calculated?

  2. Does the full system improve on simpler alternatives under equal verification budgets? Could a limited comparison use the same checkpoint, starting designs and verification calls (with only random allowed edits and/or no uncertainty bonus)? Could the authors provide the calibration curve, acceptance/rollback rates, edit distributions and predictor-agreement results described in Appendix A.14? Separately, what existing evidence supports the GRPO uncertainty penalty?

Reviewer_qJV3OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

EPIFORGE treats epitope-conditioned design as a verified repair rather than one-pass generation. A candidate variable region is initialized on a fixed framework, and a closed loop then proposes an edit, predicts its effect, verifies the most promising candidates with a real structure prediction, and either keeps or rolls back. Discrete-diffusion initializer writes CDR sequences onto a fixed frame coarse-grain graph. It is followed by four distinct sections of the pipeline. SKETCH is an instruction-tuned LLM policy which gives out a short integer genotype, GUAGE is a parameter-free decoder which projects that genotype on to a mutation set, ECHO is a sequence only surrogate to predict resulting interface metrics with its own variance and without a structure prediction and SCOUT ranks candidates by an upper confidence bound.

Critical strengths

Restricting acceptance to only a real prediction of the edited complex and deterministic metrics computed from it and rolling back anything that does not improve the state is sensible.

  • Costing verification budget instead of updating parameters for edits the surrogate cannot place is sensible.
Critical weaknesses

Manuscript inconsistencies: Three claims in the text (Section 3.5, A.7, A.6) rely on section A.14 for supporting evidence. Section A.14 says it ‘further contains’ 7 analyses but it does not. Method “AbDesign” mentioned in Figures 2, 3 and 4 does not appear in the paper (in no baseline list, no table, and no reference); its only three occurrences in the submission are these three figures. A.12 titled ‘Implementation and hyper parameters’ shares no values for important parameters optimizer settings, architecture sizes/training budget, or information on the instruction-tuned language model used in SKETCH. Figure 3 asserts All-CDR backbone RMSD 0.78 Å; “the only sub-Ångström value in the panel", but no all-CDR RMSD is shown for any method in any panel; Figure 4 likewise asserts CDR-H3 recovery for two baselines but the metric compared against in the panel is all-CDR AAR.

  • Reporting inconsistencies: The headline Table 1 and Table 7 seem to be reporting different RMSD on the same datasets. The tables have the same recovery metrics but not RMSD. Moreover, within Table 7, the three length bins have counts 23 + 209 + 112 = 344, exactly the Overall row's count, so I’d assume the bins partition the test set and so the count-weighted bin means should ideally reproduce the Overall row, which it doesn’t? The AbODE and AbDPO identical rows appropriately flagged by the authors point to an unresolved pipeline issue which calls into question other values in the table. Similarly, Table 3 Boltz-2’s DockQ and epitope recall, and ESMFold deltas match exactly to 4 decimal places, while other metrics differ, which seems implausible. Table 1’s dyMEAN and EPIFORGE tie to Protenix epitope recall also begs the question of how two independent methods agree to 4 decimals places on a mean over 344 complexes? Section 4’s experimental setup declares SKEMPI 2.0, AB-Bind, an anti-HER2 deep mutational panel and a measured affinity panel yet they are not mentioned again and no results are reported against them.

  • Single-run reporting with no seed variance.

Questions for authors
  • The CDR-H3 half of Table 1's Mean cell gives 4.178 Å (dyMEAN) and 3.735 Å (EPIFORGE); Table 7's Overall (test) row gives 6.821 Å and 4.104 Å, with the recovery columns agreeing exactly at 40.08 and 45.32. What differs between the two computations? The three length bins in Table 7 sum to 344, the Overall count. Why does the count-weighted bin mean not reproduce the Overall row (Base recovery 42.50 vs 40.08; Base RMSD 7.752 vs 6.821)? Please clarify how the calculations were made.

  • Please provide the missing analysis mentioned in A14 especially calibration curve of the ECHO variance head against realized error and the agreement between the in-loop verifier and reporting predictor on the same designs.

  • Please explain how multi-chain antibody-antigen complexes was obtained using ESMFold. Please also cite ESMFold.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.