- 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?
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.