acceptodds
Under review as a conference paper at ICLR 2027

iGC: Antibody Optimization via Reinforcement Learning over a Learned Somatic Hypermutation Prior

Abstract

Antibody affinity maturation is a natural optimization loop. Individual B cells propose mutations through somatic hypermutation (SHM), variants with higher affinity for antigen are positively selected, and the process repeats cyclically, giving rise to a population of high-affinity antibody producing B cells. A computational model of this process should therefore capture both how mutations arise and how they are selected for antigen-specific affinity. However, modeling these two components jointly is challenging: mutation models capture the sequence patterns of SHM but do not directly optimize antigen binding, while affinity optimization can improve binding without accounting for biologically plausible mutation trajectories. We introduce iGC (in silico Germinal Centers), a two-stage reinforcement learning (RL) framework for antigen-specific antibody optimization. Stage1 learns an SHM mutation prior from parent–child pairs in B-cell lineage trees and domain-adapts it to the target affinity-maturation dataset. Stage2 initializes an RL policy from this adapted prior and optimizes antigen binding while remaining Kullback–Leibler (KL) regularized toward a frozen copy of the prior. Direct optimization of affinity rewards can lead to degenerate solutions, including excessive use of a small number of highly rewarding substitutions or convergence toward structurally tolerated framework mutations. To mitigate these effects, we introduce a Reward–Tax Meta-Gradient objective. RewardNet smooths the affinity landscape, while TaxNet discourages reward overexploitation; their difference defines a learned meta-reward for policy optimization. We evaluate whether iGC achieves both objectives: preserving SHM-consistent mutation behavior and improving antigen-specific binding. The mutation prior more than doubles the strongest baseline on mutation-site prediction. Following RL optimization, iGC preserves strong SHM sequence structure, with 57.0% of predicted mutation positions falling within WRC hotspots (canonical DNA motifs preferentially targeted during SHM). iGC achieves a strong exploration–exploitation balance, yielding competitive reward–mutation-diversity products across candidate-selection cutoffs. These results demonstrate that iGC can jointly model biologically plausible mutation generation and antigen-specific affinity selection, providing an in silico framework for iterative antibody affinity maturation. Code and implementation details are available at https://anonymous.4open.science/r/iGC-0A08/README.md.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Reinforcement Learning, Antibody Optimization, B cell evolution, Adaptive Immune System

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

This paper presents iGC, a two-stage framework for antibody optimization. It first learns an SHM mutation prior from B-cell lineage data and adapts it to the target dataset. It then applies KL-regularized reinforcement learning with a Reward–Tax Meta-Gradient objective to improve antigen binding while maintaining mutation diversity. The experiments evaluate mutation prediction, cumulative DMS reward, and mutation diversity.

Critical strengths
  1. Effective use of existing data. The framework combines repertoire-derived parent–child pairs with affinity-labeled mutation trajectories. This allows abundant mutation data to support prior learning while using more limited experimental binding measurements for optimization.

  2. Biologically motivated formulation. Separating mutation generation from affinity-driven selection corresponds naturally to the diversification–selection cycle of affinity maturation. Regularizing the policy toward an SHM prior also provides a reasonable mechanism for retaining biologically plausible mutation preferences.

Critical weaknesses
  1. Limited methodological novelty. The framework primarily integrates established components. RTMG itself builds on prior intrinsic-reward meta-learning. While the application is relevant, the paper does not sufficiently establish a distinct algorithmic contribution beyond combining and adapting these techniques to antibody optimization.

  2. Risk of overfitting to the reward proxy. Although RTMG reshapes the training reward, both affinity supervision and the principal evaluation in Table 2 rely on the same DMS scoring landscape. Candidates are ranked and evaluated using cumulative DMS reward, so improvements may reflect increasingly effective exploitation of this fixed proxy rather than corresponding improvements in actual binding.

  3. Insufficient validation of additive scoring for generated multi-mutants. Section 4.2 computes trajectory rewards by summing single-mutant DMS effects, ignoring epistatic interactions. Consequently, individually beneficial substitutions may receive a favorable combined score even when their combination compromises binding. The original GCReplay study provides experimental support for additivity within this antibody system, but this does not establish its accuracy for the combinations actively selected by iGC. The evaluation therefore remains an insufficiently validated proxy for the actual affinity of novel multi-mutant candidates.

Questions for authors
  1. DMS coverage and missing-score handling. What fraction of generated substitutions has a directly available DMS score, broken down by method and mutation budget? How are substitutions absent from the lookup table handled? The released scoring script appears to assign zero to missing entries, please confirm whether this rule was used for Table 2.

  2. SHMReplay’s higher diversity and compatibility with RTMG. Table 2 shows consistently higher mutation diversity for SHMReplay than for iGC. What explains this gap, given RTMG’s explicit goal of preventing diversity collapse? Could the proposed affinity-guided optimization and Reward–Tax objective be applied to SHMReplay under matched data and generation budgets? Such a comparison would clarify whether combining SHMReplay’s broader exploration with the proposed optimization yields a better reward–diversity trade-off, and whether the gains depend on the mutation model used by iGC.

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

iGC first learns a sequence-conditioned somatic hypermutation prior and then adapts a mutation policy using logged trajectories, KL regularization toward the prior, and a Reward–Tax Meta-Gradient objective. Stage 1 evaluates mutation prediction on donor-held-out human repertoire data. Stage 2 generates multi-step VH mutation candidates in one antigen and V-gene setting. The main benchmark ranks candidates by cumulative single-mutation DMS reward and reports that reward, distinct mutation identities, and their product. The appendix also describes a structural binding-energy proxy, whose relationship to the reported benchmark needs clarification.

Critical strengths
  • The SHM prior has a measurable held-out prediction benefit. On the donor-held-out evaluation, mutation-site top-1 accuracy is 11.929%, compared with 5.307% for DASM mutation × selection; amino-acid accuracy at the true site is much closer. The separate metrics make the location of the improvement interpretable and provide substantive evidence for the prior component.
Critical weaknesses
  • The main affinity evidence does not validate the generated combinations. The objective and Table 2 evaluation add single-mutation DMS rewards, while the generated multi-mutant candidates are not directly measured for binding. This does not establish their combined effects or account for epistasis. Because improved binding of multi-step candidates is a central claim, I need independent combination measurements or a predictor validated for those combinations. The structural-proxy description in Appendix D.2 cannot resolve this concern without corresponding results and validation.

  • Mutation-identity diversity does not establish useful candidate coverage. The reported #Dist counts distinct substitution identities across a selected pool, rather than distinct complete sequences or functional candidates. Consequently, an improved DMS × #Dist product does not show that the method supplies more useful alternatives for downstream selection. The contribution of RTMG would be substantially clearer with complete-sequence coverage and diversity assessed among independently validated useful candidates.

Questions for authors
  1. What exactly produced Table 2: cumulative single-mutation DMS, the structural ΔΔG proxy, or both? Please specify the ranking score, reported score, and role of the three structural modeling runs described in Appendix D.2.
  2. Do you have independent multi-mutant binding measurements or validation of a combination-level predictor? What evidence shows that gains in the additive reward correspond to improved binding of the generated sequences?
  3. I encourage the authors to expand the existing discussion of CoSiNE[1] by distinguishing reward-shaped policy training from inference-time Guided Gillespie steering, and explaining what is learned about mutation and selection in each formulation.
  4. At matched generation budgets, how many distinct complete sequences and independently supported useful candidates does each method produce? Does RTMG’s advantage remain under those measures? : )

[1] Lu, S. Z., Vermani, A., Sanno, K., Lu, J., Matsen IV, F. A., Jagota, M., & Song, Y. S. (2026). Conditionally Site-Independent Neural Evolution of Antibody Sequences. ArXiv, arXiv-2602.

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

The paper introduces iGC for antibody optimization. A Transformer learns a somatic hypermutation prior from lineage pairs and is adapted to GCReplay. The policy is then optimized using DMS-derived rewards, KL regularization, and a learned reward–tax objective intended to preserve mutation diversity. Experiments report improved mutation-site prediction and favorable trade-offs between cumulative single-mutant DMS scores and mutation diversity, alongside SHM motif analysis and held-out sequence rediscovery.

Critical strengths
  1. Stage 1 substantially improves mutation-site prediction on held-out donors. Table 1 reports pos@1 of 11.929%, compared with 5.307% for the strongest baseline. This is a clear empirical contribution.

  2. The reward-shaping approach shows a useful improvement over directly optimizing raw rewards. In Table 2, Top-100 mean DMS increases from 0.844 to 1.010, while the number of distinct mutations increases from 92.6 to 149.4.

Critical weaknesses
  1. The data split is inconsistent with the construction of the training signal. Appendix E.1 builds the sequence-frequency score from “all GCReplay lineage trees,” and this score enters both TaxNet and the RewardNet objective. This conflicts with the claim in Appendix A.2 that test trees are unseen during training. As written, held-out sequence statistics influence training, so I think the rediscovery results cannot yet establish the claimed generalization.

  2. The claimed structural validation is not supported by the reported results. Section 5.2 and Table 2 evaluate cumulative single-mutant DMS scores, whereas Appendix D.2 describes three structural modeling runs per candidate and binding-energy evaluation. No corresponding structural results are presented, although the conclusion explicitly refers to them. The demonstrated improvement therefore remains limited to the additive DMS scoring objective.

Questions for authors
  1. Were all sequence statistics used during training computed exclusively from the training trees in each split?

  2. Where are the structural evaluation results described in Appendix, and do they agree with the DMS-based comparisons?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.