acceptodds
Under review as a conference paper at ICLR 2027

MGPO: Mutant Group Policy Optimization for StructureDiverse Protein Reinforcement Learning

Abstract

Proteins can admit multiple structurally distinct solutions for the same amino-acid sequence, making broad structural exploration important for protein generative models. Mutation offers a natural way to expand policy exploration beyond ordinary autoregressive sampling, but we identify a previously under-characterized failure mode of mutation-augmented reinforcement learning: evolutionary exploration produces heterogeneous structural neighborhoods with different reward offsets, so a single global baseline can assign negative advantages to mutants that are improving within their own neighborhood. We introduce Mutant Group Policy Optimization (MGPO), which couples structured evolutionary exploration with neighborhood-aware credit assignment in the protein-structure-token space. MGPO generates structurally diverse mutant populations through reward-blind mutation and quality-diversity search, organizes them using reward-independent structural geometry, and computes each mutant's advantage relative to its full structural neighborhood before representative compression. This preserves learning signals for promising mutant families that would otherwise be suppressed by global population statistics. On the CAMEO22 benchmark, MGPO substantially improves mean lDDT-CA and TM-score. Diversity is evaluated in both structure token space and decoded 3D space; the measurements show that MGPO can improve the diversity of generated protein structures. A matched same-layout comparison, together with local-only ablations, shows that neighborhood-local credit assignment accounts for most of the improvement, while the evolutionary search provides the diverse structural neighborhoods on which this mechanism operates. These results establish MGPO as a mutation-aware policy-optimization framework for jointly improving structural quality and maintaining population-level exploration.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein Language Model, Reinforcement Learning, Protein Folding

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

MGPO fine-tunes an autoregressive protein-structure generator using evolutionary mutants. It partitions mutants into reward-blind structural neighborhoods and computes advantages relative to each neighborhood’s mean reward. The motivation is that global centering can discourage candidates that outperform their local neighborhood. On 183 CAMEO22 targets, the paper reports improvements in native-reference quality and population diversity under policy-only evaluation.

Critical strengths

An interesting attempt to combine autoregressive protein-structure generation with mutation-augmented RL. The paper identifies a real issue in mutation-augmented RL: when exploration spreads samples across neighborhoods with different reward offsets, a global baseline suppresses candidates that are improving within their own group. The remedy is clean, and computing the group baseline on the full cluster before representative compression is a careful detail.

Critical weaknesses

W1. The problem formulation and modeling choices are insufficiently motivated. The paper shifts from autoregressive structure prediction to credit assignment in mutation-augmented RL without clearly connecting these goals. Although structure tokens follow residue order, they have no intrinsic left-to-right causal ordering, so the suitability of this factorization requires justification. The choice of Qwen over a protein-specific model is also unclear. Given the subsequent 200B-token protein-domain training, it is unclear what benefit its human-language pretraining provides over a protein-specific initialization or training the same architecture from scratch. Furthermore, the absence of direct comparisons with established predictors such as AlphaFold2, ESMFold, or DPLM-2 leaves the approach’s practical competitiveness unresolved.

W2. The necessity of RL and the validity of the resulting diversity remain unclear. Since reference structure tokens are available and token accuracy contributes to the reward, continued reference-based SFT is an important missing baseline. Moreover, increased structural diversity may reflect prediction errors or physically implausible structures rather than meaningful conformational variation.

Questions for authors

Q1. Why is a Qwen-based autoregressive model preferable to directly using or fine-tuning DPLM-2? Q2. How does MGPO compare with continued SFT on the reference structure tokens using the same training data and a comparable compute budget? Q3. How do decoded mutants compare with policy-sampled anchors in terms of physical validity (e.g., MolProbity scores, clash rates, and bond-length statistics)? How do AlphaFold2, ESMFold, and DPLM-2 perform on CAMEO22 under the same evaluation protocol?

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

This paper proposes mutant group policy optimization for structure diverse protein reinforcement learning; the mutations are generated with quality and diversity, and later computed with advantage relative to its full structural neighborhood, which helps to preserve important signals.

Critical strengths

The idea of using mutants to conduct group search to expand the reinforcement learning search space is interesting, and it connects domain specific knowledge with real-world problems and applications.

Critical weaknesses

The paper is hard to read and follow, clearly the figures and tables in the paper are generated with AI and are not human polished, the font in the figure is extremely small, and hard to be understood. While the authors claimed that their MGPO gives ‘substantial’ improvement, the gain is within 2-3% compared to SFT baselines. Given that the experiments are only done on widely known datasets, the gain is very minimal, and lack comparisons with other approaches. Besides, the MGPO name is also quite arbitrary, actually there has been a lot of papers on different ‘MGPO’ stuff. No demonstrations / Figures of protein structures are provided in the manualscript, nor did the paper show any real development for this method to be applicable in real-world protein synthesis / folding or other applications. The RL algorithm the paper introduced itself is not novel, the only novelty is about how the exploration of space changes, so it is hardly interesting by introducing it as a ‘new RL algorithm’.

Questions for authors

None

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

The paper proposes MGPO for reinforcement learning of autoregressive protein structure generation. It combines evolutionary mutation in discrete structure-token space with structure-based grouping and neighborhood-local advantage estimation. The central motivation is that a global reward baseline can penalize candidates that improve within lower-reward structural neighborhoods. Experiments on CAMEO22 report improved average structural accuracy and sample diversity. Same-layout controls and fixed-batch interventions support the contribution of local credit assignment.

Critical strengths
  1. The same-layout global-baseline control and fixed-batch interventions provide useful evidence that local credit assignment changes learning behavior and contributes to the reported gains.

  2. Evaluation uses a fixed number of policy-only samples and reports average accuracy alongside diversity. The improvements therefore cannot be explained simply by increasing test-time sampling and selecting the best structure.

Critical weaknesses
  1. The biological significance of the increased diversity remains unclear. Pairwise structural distance does not distinguish meaningful alternative conformations from prediction errors or nonphysical variation. Although the paper acknowledges this limitation, the absence of known multi-state or ensemble-level validation limits the significance of its structure-diversity contribution.

  2. Local improvement is not necessarily evidence that a structural neighborhood is worth preserving. A candidate can receive positive local advantage while remaining globally poor. The sign-reversal analysis establishes different update directions, but does not independently establish that the affected candidates represent useful exploration. Stronger evidence connecting preserved neighborhoods to subsequent quality gains would strengthen the central interpretation.

Questions for authors
  1. Can you provide evidence that the additional diversity covers experimentally observed alternative states or physically plausible structural changes, rather than primarily reflecting prediction variation? This would materially affect my assessment of the biological significance.

  2. Do candidates receiving positive local but negative global advantages subsequently produce higher-quality descendants? Can you quantify whether local credit assignment preserves productive neighborhoods versus persistently low-quality ones?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

We sincerely thank all reviewers for their time and constructive feedback. We are encouraged that reviewers recognize the core issue we identify — that a global baseline in mutation-augmented RL can suppress candidates improving within their own structural neighborhood — and find our remedy "clean" with "careful details" (yZMX), the same-layout controls "useful evidence" (cCpJ), and the group-search idea "interesting" with a real-world connection (YyTe). Below we address the main concerns.

1. Biological significance and validity of diversity (cCpJ-W1, yZMX-W2). We agree this is the most important open question. Our current diversity metrics operate in structure-token and decoded 3D space and do not distinguish experimentally observed alternative conformations from prediction variation. Validating against multi-state ensembles, physical-plausibility checks (e.g., MolProbity-style statistics, clash rates, bond geometry), and known alternative-state benchmarks is a clear next step, and we will pursue it in follow-up work.

2. Do preserved neighborhoods produce quality gains? (cCpJ-W2). We acknowledge that positive local advantage alone does not prove a neighborhood is worth preserving. Our sign-reversal analysis shows that local credit assignment changes the update direction, but we agree that longer-horizon evidence — tracking whether locally-advantaged candidates yield higher-quality descendants — would strengthen the interpretation. We will investigate this lineage-level analysis in future work.

3. Baselines and comparisons (YyTe, yZMX-W1/W2). We will clarify in the revision that our contribution is the credit-assignment mechanism and evaluation protocol rather than a new RL algorithm, and RL component is ablated against same-layout global-baseline controls, which isolates its effect. We will also temper "substantial" to describe the 2–3% relative gains precisely, with exact numbers and variance.

4. Presentation (YyTe). We apologize for the readability of figures and tables. All figures will be re-rendered with legible fonts and human-polished formatting, and we will add example decoded protein structures for qualitative inspection.

5. Naming (YyTe). We were unaware of the name collision with prior "MGPO" usage and will rename the method or add disambiguating context.

Limitations and future work. We acknowledge that connections to established predictors (AlphaFold2, ESMFold, DPLM-2), reference-token SFT baselines, physical-validity evaluation, and ensemble-level validation are not included in the current submission. Addressing these requires substantial additional experimentation beyond the rebuttal window; we commit to pursuing these directions and reporting them in an extended future version of this work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.