acceptodds
Under review as a conference paper at ICLR 2027

Euterpe: Efficient Multi-Level Preservation of Pretrained Performance during Preference Alignment for Protein Inverse Folding

Abstract

Protein inverse folding (IF) aims to design amino acid sequences for a given protein backbone, making it a fundamental tool for protein design and engineering. Recent Direct Preference Optimization (DPO)-based IF post-training improves developability properties such as solubility and thermostability, but can drive the aligned model away from the pretrained IF model. Existing multi-objective methods address this trade-off by incorporating costly folding-derived structural signals as additional designability alignment objectives. Our analysis shows, however, that the resulting structural gains largely recover the performance lost during developability alignment, rather than improving beyond the pretrained IF model. This raises a natural question: can we regulate developability alignment using the pretrained IF model itself, without costly folding-derived feedback? We introduce Euterpe, an efficient semi-online preference optimization framework that uses the pretrained IF model as a reference prior to regulate alignment. Pair-level Reweighting uses reference-model energies to downweight preference pairs weakly supported by the pretrained distribution, while Token-Level Thinking-Pattern-Guided Reweighting modulates residue-wise updates according to the thinking pattern induced by the pretrained IF model. Built on ProteinMPNN, Euterpe preserves designability and AAR at levels comparable to methods using explicit folding-derived structural feedback, while eliminating such feedback during alignment, requiring fewer GFLOPs, and achieving up to a speedup.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Inverse Folding, Post-training, Direct Preference Optimization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_URrbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Euterpe proposes an economical semi-online preference optimization framework to align pretrained ProteinMPNN toward protein developability properties while preserving inverse-folding performance. Instead of relying on costly folding-derived structural feedback as in MoMPNN, Euterpe uses signals from the frozen pretrained ProteinMPNN: a pair-level weight based on group-standardized reference sequence energy and token-level weights based on Top-K amino-acid support. On CATH 4.3, Euterpe matches MoMPNN closely in solubility, thermostability, designability proxies, and AAR, achieving up to a 6.9x per-epoch speedup in single-property settings and reducing to ~1.7x when evolutionary plausibility is included. The paper also provides global KL comparisons, component ablations, CASP15 TS45 transfer results, and an independent solubility evaluation via SoluProt.

Critical strengths
  1. This paper has strong practical motivation and offers clear computational savings. Improving solubility or thermostability without degrading structural recovery is an important challenge in inverse-folding post-training. The detailed runtime and GFLOPs breakdown clearly shows that replacing repeated ESMFold evaluations with reference-model likelihood queries substantially reduces post-training overhead in single-property settings.

  2. This paper includes informative empirical comparisons that help contextualize the reported gains. Beyond comparing against MoMPNN on CATH 4.3, the authors benchmark against β\beta-tuned DPO and explicit token-level KL regularization, analyze optimization dynamics, test zero-shot transfer on CASP15 TS45, and validate solubility with an independent predictor. These experiments support the practical feasibility of reference-guided preservation.

Critical weaknesses
  1. The distinctive methodological contribution is not adequately isolated, and the pair-level module shows marginal utility. Preference-pair and token-wise DPO reweighting have clear precedents (e.g., WPO, TIS-DPO). Euterpe introduces nine heuristic hyperparameters (τΔ,τe,η,λ,Kw,Kl,α,γ,ϵ\tau_\Delta, \tau_e, \eta, \lambda, K_w, K_l, \alpha, \gamma, \epsilon) whose thresholds are retuned across tasks. Moreover, Figure 5(B) and Figure 8 show that token-level reweighting accounts for nearly all preservation gains, while removing pair-level reweighting or setting η=1.0\eta=1.0 barely affects designability. At the token level, combining Top-K indicators from ywy_w and yly_l into a shared weight gig_i ignores that ywy_w and yly_l are evaluated under different decoding orders (ow≠olo_w \neq o_l) and prefixes. The paper lacks controlled baselines using continuous reference probabilities, shuffled token weights, or simply plugging reference energy directly into MoMPNN (Eq. 4) in place of TM-score.

  2. The central proxy assumption and broader generalization are insufficiently supported. High pretrained likelihood does not guarantee recovery of the target fold, especially on novel or out-of-distribution backbones, and this correlation is never quantified directly. Empirically, the evaluation omits the de novo and binder design settings studied in MoMPNN; on CASP15 TS45 (Table 7), Euterpe lags behind MoMPNN in pLDDT and RMSD < 3Å. Furthermore, thermostability is only evaluated by the training predictor (TemBERTure), Table 4 reports variance across five inference resamplings rather than independent training seeds, and the 6.9x speedup stems entirely from dropping ESMFold (shrinking to ~1.7x when EP is included in Table 2).

Questions for authors
  1. Under the same training protocol, how does Euterpe compare with simpler alternatives to the proposed reweighting scheme, such as continuous reference-probability weighting, separately weighting the winner and loser branches under their own decoding contexts, or shuffled token weights with matched overall strength? It would also be helpful to compare against directly using the reference energy −Eref-E_{\text{ref}} in MoMPNN’s multi-objective margin instead of TM-score.

  2. How reliable is the frozen ProteinMPNN energy as a proxy for structural quality, especially for low-likelihood sequences and unfamiliar or OOD backbones? Could the authors report its correlation with ESMFold TM-score and RMSD across different energy ranges and provide examples where the proxy fails?

  3. Table 2 suggests that most of the reported 6.9× speedup comes from removing ESMFold, while the advantage is closer to 1.7× when EP is included. Under a matched compute budget and comparable target-property performance, how does Euterpe compare with simpler no-folding baselines such as tuned or early-stopped DPO+KL on the structure–property trade-off?

  4. Could the authors evaluate Euterpe on de novo backbones, as done in MoMPNN, and validate thermostability with an independent predictor? It would also be useful to report variation across independent training runs rather than only inference resampling, and to explain the drop in pLDDT and RMSD<3Å performance on CASP15 TS45 shown in Table 7.

2 comments
AuthorsOpenReview ↗

Response to Reviewer URrb (1/2)

Thanks for your careful reviews!

The complementary role of pair-level and token-level reweighting.

Answer: We agree that token-level reweighting contributes most of the preservation gains. However, pair-level reweighting provides complementary sequence-level information by capturing reference-ranking conflicts and overall winner support, which cannot be fully characterized by local Top-K indicators. As shown in Fig. 5(B), combining both components achieves better preservation than token-level reweighting alone, demonstrating the benefit of jointly leveraging sequence-level and residue-level reference information.

Different decoding contexts for winner and loser sequences.

Answer: We would like to clarify that Euterpe explicitly preserves the original decoding order sampled for each sequence during rollout. The winner and loser are evaluated separately under their respective decoding orders and autoregressive prefixes, as described in Section 4.3 and Eq. 8.

The resulting indicators characterize whether each residue is supported by the frozen reference model under its own generation context. They are subsequently combined to determine the relative contribution of the corresponding residue-wise preference signal. Therefore, our formulation does not assume that the winner and loser are evaluated under identical conditional distributions, nor does it directly compare their conditional probabilities as though they shared the same prefix. Instead, it combines their respective local support indicators to regulate preference optimization.

Additional weighting baselines and methodological novelty

Answer: Euterpe's contribution goes beyond specific reweighting formulations: it demonstrates that pretrained IF knowledge can replace costly folding-derived feedback for preserving existing capabilities during developability alignment. We have already compared against property-only DPO, β-tuned DPO, explicit KL regularization, and component-wise ablations.

The suggested alternatives represent different possible design choices rather than established baselines with uniquely defined formulations. Our contribution goes beyond optimizing a particular weighting function, and we have already provided comparisons against property-only DPO, β-tuned DPO, explicit KL regularization, and component-wise ablations. While these alternatives may provide additional insights, our central contribution is to demonstrate effective preservation of pretrained IF capabilities without costly folding-derived feedback, rather than to establish the optimal weighting formulation among all possible variants.

Hyperparameter sensitivity and Generalization Robustness

Answer: We acknowledge that Euterpe introduces multiple hyperparameters through its two-level reweighting scheme. However, a larger number of hyperparameters does not necessarily imply a need for fine-grained tuning. Our sensitivity analyses in Appendix C.1 (Figs.~6--8) show generally stable performance across the tested configurations, with parameter changes primarily controlling the trade-off between developability alignment and preservation of pretrained IF performance.

For token-level Top-KK, TM-score, pLDDT, and AAR remain within 0.7290.729--0.7450.745, 77.5777.57--78.3378.33, and 0.3690.369--0.3840.384, respectively, across all nine configurations. Lower solubility occurs mainly under the more conservative Kw=1K_w=1 configurations, while the remaining six configurations achieve solubility scores of 0.8700.870--0.9260.926. For pair-level thresholds, solubility, TM-score, pLDDT, and AAR vary only within 0.8850.885--0.8980.898, 0.7360.736--0.7420.742, 77.8477.84--78.1778.17, and 0.3740.374--0.3780.378, respectively. The attenuation-coefficient analysis likewise shows stable structural preservation.

These results indicate that effective preservation is achieved across multiple tested configurations, rather than depending on a narrowly tuned hyperparameter combination. Moreover, most parameters are shared between the two single-property tasks in Table~5, and the same CATH checkpoints and hyperparameter configurations are directly transferred to CASP15 without additional tuning, further supporting their cross-benchmark applicability.

AuthorsOpenReview ↗

Response to Reviewer URrb (2/2)

Reliability of Pretrained Reference Energy as a Structural Proxy

Answer: We would like to clarify that Euterpe does not assume reference likelihood is equivalent to structural designability, nor that all structurally compatible sequences must receive high reference support. Our objective is to improve developability while preserving the designability and AAR already achieved by pretrained ProteinMPNN. The frozen IF model therefore serves as a preservation prior, rather than a predictor of structural quality.

Importantly, Euterpe employs soft reweighting rather than hard filtering, while semi-online rollouts continue exploring under the current policy. Thus, low-reference-support but structurally compatible sequences are not excluded, and their existence does not contradict our objective.

While correlations between reference energy and folding metrics could further characterize the prior, our method does not require reference likelihood to be a calibrated structural-quality predictor. Its effectiveness is instead evaluated directly through the preservation of designability and AAR during developability alignment.

Computational Efficiency and Simpler DPO/KL Baselines

Answer: We agree that the primary speedup comes from eliminating repeated ESMFold evaluations. This is precisely the computational bottleneck Euterpe aims to eliminate. Table 2 transparently reports the full cost breakdown, including the 6.9× speedup for single-property alignment and the approximately 1.7× speedup when EP evaluation dominates the computational cost. Our contribution is to achieve comparable property–designability trade-offs without expensive folding-derived feedback, rather than to accelerate components unrelated to structural preservation.

Besides, we have already investigated simpler alternatives in Fig. 4, including β-tuned DPO and DPO with different KL regularization strengths (equivalent to MoDPO+KL in the single-objective setting). At comparable solubility, Euterpe achieves better TM-score and RMSD-based preservation while maintaining comparable pLDDT, demonstrating a more favorable structure–property trade-off. Furthermore, Fig. 5(A) shows that even when property-only DPO surpasses Euterpe's final solubility at an early epoch, its structural preservation remains inferior. These results support the benefits of our reference-guided preservation beyond conventional KL regularization or early stopping.

CASP15 generalization

Answer: We acknowledge that Euterpe achieves lower pLDDT and RMSD < 3Å success rates than MoMPNN on CASP15 TS45. However, the comparison is mixed across metrics. Euterpe achieves higher TM-score (0.5904 vs. 0.5875), RMSD < 2Å success rate (0.3111 vs. 0.2815), and AAR (0.3591 vs. 0.3578), with nearly identical solubility (0.8537 vs. 0.8555). Euterpe also achieves higher solubility on 29 of the 45 targets.

Importantly, these results are obtained by directly transferring the same checkpoints and hyperparameters from CATH 4.3 to CASP15 TS45, without additional training or tuning. We do not claim uniform superiority across every metric, rather, these results support a comparable overall property–designability trade-off without folding-derived feedback.

Additional Evaluation and Limitations

Answer: Our study focuses on preserving pretrained IF capabilities during developability alignment, with extensive evaluation on held-out CATH 4.3 structures and zero-shot transfer to CASP15 TS45 without additional training or tuning. We follow the same evaluation protocol as MoMPNN for fair comparison and further validate solubility using the independent SoluProt predictor. These results support our conclusions within the intended scope. We will further clarify the distinction between inference-time resampling variability and independent training-seed variability, explicitly state the current evaluation scope, and discuss de novo/binder design and independent thermostability validation as future research directions in the revised manuscript.

We hope our responses have addressed your concerns, thank you!

Reviewer_xs6xOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Euterpe, a semi-online preference optimization framework for protein inverse folding. By employing a fine-grained reweighting strategy, Euterpe replaces external folding models with the inverse folding model itself as a reference prior. As a result, Euterpe successfully preserves the structural knowledge of the pretrained IF model while remaining capable of learning extra functional protein features, such as solubility and thermostability.

Critical strengths
  1. The core motivation is highly compelling and practical. Incorporating online structural feedback introduces prohibitive computational bottlenecks. By demonstrating that the structural preservation signal can be largely extracted from the pretrained inverse folding model itself, Euterpe successfully cuts down the training overhead, achieving up to a 6.9×6.9\times speedup without sacrificing core target properties.
  2. Unlike standard property-only DPO, which rapidly degrades structural compatibility and native recovery rates, Euterpe successfully prevents catastrophic forgetting of the backbone geometry.
Critical weaknesses
  1. While replacing dynamic folding feedback with the static reference model saves computation, it inherently limits the model's exploratory boundaries. By heavily penalizing or downweighting pairs and tokens that fall outside the pretrained model's high-support regions (e.g., via the reference support threshold τe\tau_e and Top-KK filtering), the method might discard potentially valuable evolutionary trajectories. For example, novel sequence variations that exhibit low sequence similarity to the reference (low ref-support) but could still achieve high structural compatibility.
  2. Although Euterpe successfully replaces the folding model with the IF model, it also introduces extra hyperparameters. Such as pair-level thresholds (τΔ,τe\tau_\Delta, \tau_e), attenuation factors (η,λ\eta, \lambda), and token-level Top-KK sizes (Kw,KlK_w, K_l) alongside state-modifying weights (α,γ,ε\alpha, \gamma, \varepsilon). As shown in the appendix sensitivity analyses (e.g., Fig. 6 & Fig. 7), the model performance is highly sensitive to these hyperparameters (e.g., varying the Top-KK setting shifts the solubility from 0.799 to 0.926, while trade-off metrics fluctuate significantly). This raises serious concerns regarding hyperparameter tuning costs and generalization robustness when transferring Euterpe to entirely different protein design tasks.
Questions for authors

See weakness.

3 comments
AuthorsOpenReview ↗

Response to Reviewer xs6x

Thanks for your careful reviews!

Limited exploration due to reliance on pretrained reference support

Answer: We agree that low likelihood under a pretrained IF model does not necessarily imply structural incompatibility. However, Euterpe uses reference information to regulate preference updates, rather than constrain the sequence exploration space. Notably, our reference energy measures structure-conditioned sequence likelihood relative to other candidates generated for the same backbone, rather than sequence similarity to a reference sequence.

Moreover, Euterpe's semi-online strategy helps maintain candidate diversity by resampling sequences from the evolving policy at a relatively high sampling temperature (T=1.0T=1.0), rather than relying on a fixed offline candidate pool. After sampling, reference support is used for soft reweighting rather than hard filtering. Pair-level weights attenuate preference contributions, while all token-level weights remain strictly positive (1,α,γ,ϵ>01,\alpha,\gamma,\epsilon>0). Thus, low-support candidates are not discarded solely due to their reference support. The property-score threshold δ\delta only filters pairs with insufficient target-property differences.

We acknowledge that reference-guided reweighting may reduce the contribution of some potentially valuable, weakly supported variants. Our goal, however, is to balance exploration with the preservation of pretrained structural knowledge, not to restrict exploration to the pretrained model's original distribution.

Hyperparameter sensitivity and generalization robustness

Answer: We respectfully disagree with the characterization that Euterpe exhibits high hyperparameter sensitivity. Our robustness experiments in Appendix C.1 (Figs. 6–8) demonstrate that the hyperparameters provide predictable control over the trade-off between developability alignment and preservation of pretrained IF performance.

(1) Token-level Top-K (Fig. 6). The relatively lower solubility scores are observed only when Kw=1K_w=1 (0.799–0.821), corresponding to particularly conservative configurations that place greater emphasis on preserving pretrained IF capabilities. In contrast, all remaining six configurations achieve consistently high solubility scores (0.870–0.926), including our default setting (Kw=Kl=3K_w=K_l=3, 0.894). Meanwhile, structural metrics remain consistently strong across all nine configurations, with TM-score ranging from 0.729 to 0.745, pLDDT from 77.57 to 78.33, and AAR from 0.369 to 0.384. These results demonstrate that Euterpe performs robustly across a broad range of Top-K configurations, without requiring narrowly tuned hyperparameters.

(2) Pair-level thresholds (Fig. 7). Performance varies only modestly across the tested threshold configurations. Solubility ranges from 0.885 to 0.898, TM-score from 0.736 to 0.742, pLDDT from 77.84 to 78.17, and AAR from 0.374 to 0.378. These results directly demonstrate that Euterpe is not highly sensitive to neighboring threshold choices.

(3) Pair-level attenuation coefficient (Fig. 8). Across η∈0.90,0.95,1.00\eta\in{0.90,0.95,1.00}, TM-score, pLDDT, and AAR remain particularly stable, with only minor variations. The gradual changes in solubility and structural metrics follow the expected direction of the alignment-preservation trade-off. The token-level state weights are likewise fixed positive attenuation coefficients, rather than hard selection masks or per-sequence tuning variables.

(4) Cross-benchmark generalization (Appendix C.2). More importantly, we have already evaluated Euterpe on the independent CASP15 TS45 benchmark using exactly the same checkpoints and hyperparameter configurations as in CATH 4.3, without additional training or tuning. Euterpe achieves an average solubility of 0.8537, comparable to MoMPNN's 0.8555, and obtains higher solubility on 29 out of 45 targets. It also achieves a higher average TM-score (0.5904 vs. 0.5875) and RMSD < 2 Å success rate (0.3111 vs. 0.2815), demonstrating a comparable overall property–designability trade-off on an independent benchmark.

We hope our responses have addressed your concerns, thank you!

After reading the authors’ responses and the other reviewers’ comments, I still have some concerns about replacing folding-derived feedback with the pretrained inverse-folding prior. I understand the distinction between soft reweighting and hard filtering, but the reported structural metrics do not fully resolve my concern about effective exploration. It would be helpful if the authors could provide additional evidence on this aspect. For example, sequence diversity among candidates satisfying a predefined TM-score threshold (such as 0.7) could help clarify whether comparable structural quality is maintained at the expense of exploring a narrower range of sequences. Regarding the multiple hyperparameters, I acknowledge the sensitivity analyses and the CASP15 evaluation without retuning. However, their transferability across models remains unclear. Additional validation using the same hyperparameters on another folding model would help strengthen the robustness claim.

AuthorsOpenReview ↗

Response to Reviewer xs6x

We thank the reviewer for the additional comments and suggestions raised in this follow-up review.

  1. We respectfully believe that the concern about exploration should be considered in the context of the objective and sampling behavior of preference optimization. DPO is designed to shift the pretrained policy toward preference-favored regions of the sequence space, rather than preserve the full diversity of the original distribution. Such distributional concentration is therefore an expected consequence of preference optimization, not necessarily evidence of restricted exploration. Importantly, Fig. 2 also compares standard DPO with the pretrained ProteinMPNN. The diversity gap between the compared methods is apparent at higher sampling temperatures but becomes negligible at lower temperatures, indicating that standard DPO does not necessarily incur substantial diversity loss under lower-temperature sampling. Euterpe further employs soft reweighting to attenuate preference-driven updates, rather than applying the unattenuated updates of standard DPO. Therefore, the shift in sequence diversity induced by Euterpe is expected to be less pronounced than that induced by standard DPO relative to the pretrained ProteinMPNN.

  2. We thank the reviewer for this constructive suggestion. We would like to clarify that the scope of this work is explicitly focused on preference optimization of pretrained ProteinMPNN, as stated in the manuscript. Accordingly, our robustness claim concerns the stability of Euterpe across evaluation benchmarks, rather than the transferability of its hyperparameters across different model architectures. In this regard, our CASP15 evaluation provides additional evidence of benchmark-level robustness: we apply the same pretrained ProteinMPNN checkpoint and hyperparameters used for CATH 4.3, without any additional tuning. That said, we agree that evaluating Euterpe across different inverse-folding architectures would be an interesting and valuable direction. Our framework is not intrinsically tied to a particular base-model architecture, suggesting potential applicability beyond ProteinMPNN. We appreciate the reviewer for highlighting this promising direction, and believe that systematically investigating cross-model generalizability and hyperparameter transferability would be a worthwhile avenue for future work.

Reviewer_Xx34OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the multi-objective preference alignment of pre-trained inverse folding (IF) models, aiming to satisfy diverse developability objectives while maintaining structural fidelity, i.e., designability. It proposes an improved strategy over the previous work MoMPNN, which avoids using a costly folding module to calculate designability alignment objective and instead employs the pretrained IF model as a reference prior to regulate alignment to maintain designability. Experiments are conducted to demonstrate lower alignment cost of the proposed method.

Critical strengths
  1. The paper is clearly written and easy to follow.
  2. The pair-level and token-level reweighting based on the pretrained IF reference signal is reasonable designed.
  3. The experiments show the significant speedup compared to MoMPNN.
Critical weaknesses
  1. The motivation for avoiding folding modules in protein design tasks may not reflect a practical challenge in real-world applications. In current protein design pipelines, structure prediction for designed sequences is typically a standard and necessary step that cannot be bypassed. Furthermore, several existing studies have already explored efficient protein folding methods, potentially reducing the computational burden associated with this step.
  2. Using a pretrained reference prior for pair-level and token-level reweighting has already been explored in existing studies [1,2]. Therefore, the technique itself is not totally novel; though I acknowledge its specific technical contribution in the context of protein inverse folding models.
  3. Some related work is missing, e.g., [3,4].
  4. Some important baselines to utilize the reference prior are not considered in the work. (1) Adding an explicit KL(πθ∣∣πref)KL(\pi_\theta||\pi_ref) as regularization for MoDPO objective; (2) adding the pretrained loss as the regularization for MoDPO objective as indicated in [3]. Also, it is worthwhile to discuss the advantages of the proposed method compared to these baselines.

Typos

  1. Figure 1, pLDD should be pLDDT.

References

[1] Yuan, Suqin, et al. "Mitigating mismatch within reference-based preference optimization." arXiv preprint arXiv:2602.11902 (2026).
[2] Christopoulou, Fenia, et al. "Sparsepo: Controlling preference alignment of llms via sparse token masks." arXiv preprint arXiv:2410.05102 (2024).
[3] Gasser, Hans-Christof, et al. "Tuning ProteinMPNN to reduce protein visibility via MHC Class I through direct preference optimization." Protein Engineering, Design and Selection 38 (2025): gzaf003.
[4] Tamatgar, Nilufer, et al. "From Structure to Function: Preference Alignment for Function-aware Protein Inverse Folding." Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 2026.

Questions for authors
  1. Doubt about the motivation for avoiding folding modules in protein design tasks may not reflect a practical challenge in real-world applications.
  2. Missing important baselines and related discussion.
7 comments
AuthorsOpenReview ↗

Response to Reviewer Xx34

Thanks for your careful reviews!

Doubt about practical relevance and motivation: Why avoid folding modules, especially as faster and more accurate folding predictors emerge?

Answer: Our goal is not to eliminate folding models from the design pipeline, but to investigate whether explicit folding-derived feedback is necessary during preference-based post-training. Accordingly, we still employ folding-derived metrics for evaluation, while avoiding repeated folding predictions during alignment. We agree that folding prediction remains essential for downstream protein design and validation.

Our motivation is empirical, not merely computational. Folding-only DPO yields limited gains over pretrained ProteinMPNN, resembling implicit temperature reduction (Fig. 2), while folding preferences help preserve designability during multi-objective alignment (Fig. 5). These observations motivated us to investigate whether such preservation is possible without explicit folding feedback. Euterpe achieves designability comparable to MoMPNN without folding-derived preference supervision. ** The resulting efficiency improvement is therefore a natural benefit of our approach, rather than its sole motivation. ** Faster and more accurate folding predictors may reduce the cost of folding feedback, but they do not change the fundamental question our work addresses: whether such feedback is necessary during DPO in the first place.

Some related work is missing

Answer: Thank you for pointing out these relevant studies. We will incorporate both references and their corresponding discussions into Sec. 2 in the subsequent revision of the manuscript.Specifically,[3] apply DPO to optimize MHC-I immune visibility, while [4] leverage external biological information to construct function-aware preference pairs and pretrained IF likelihood for hard-negative selection. In contrast, Euterpe employs reference-guided soft reweighting at both the pair and token levels to preserve pretrained inverse-folding performance during developability alignment.

Technically novel in protein IF, but not entirely novel in broader AI.

Answer: We appreciate the reviewer's recognition of our technical contribution to protein IF. While related ideas of pair- and token-level reweighting have appeared in prior work [1,2], our contribution goes beyond adopting these general ideas in a new domain. Our motivation stems from an important empirical observation that folding-derived only DPO provides limited gains over the pretrained model(Fig. 2), while in multi-objective alignment, it primarily helps prevent the degradation of pretrained designability (Fig. 5). Given the additional cost of obtaining folding-derived preferences, this raises an important question: Can we achieve effective preference alignment while preserving the capabilities of the pretrained IF model, without relying on explicit folding feedback? To this end, we develop tailored pair- and token-level mechanisms that leverage information already available in the pretrained IF model. Thus, our contributions lie not only in the specific algorithmic designs, but also in identifying and addressing this previously underexplored aspect of folding-based preference alignment.

Besides, we will also include [1,2] and further discuss their related pair- and token-level reweighting ideas in the Appendix, clarifying the distinctions from our specific formulations.

Missing baselines

Answer: Regarding the KL baseline, MoDPO+KL reduces to DPO+KL in the single-objective setting with w=1w=1. We have already evaluated this setting with different KL weights in Fig. 4. We achieve a more favorable Pareto trade-off across most evaluated metrics, while maintaining comparable pLDDT. Although DPO+KL yields slightly higher AAR under strong regularization, this is expected because explicit KL directly constrains the aligned policy toward the pretrained ProteinMPNN distribution, thereby favoring native-sequence recovery. We have already discussed and analyzed this trade-off in the main text.

[1] Yuan, Suqin, et al. "Mitigating mismatch within reference-based preference optimization." arXiv preprint arXiv:2602.11902 (2026).

[2] Christopoulou, Fenia, et al. "Sparsepo: Controlling preference alignment of llms via sparse token masks." arXiv preprint arXiv:2410.05102 (2024).

[3] Gasser, Hans-Christof, et al. "Tuning ProteinMPNN to reduce protein visibility via MHC Class I through direct preference optimization." Protein Engineering, Design and Selection 38 (2025): gzaf003.

[4] Tamatgar, Nilufer, et al. "From Structure to Function: Preference Alignment for Function-aware Protein Inverse Folding." Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 2026.

We hope our responses have addressed your concerns, thank you!

I thank the authors for their detailed responses. But my concerns have not been fully addressed.

I understand the clarified empirical motivation. It is somewhat intuitive that the model can maintain designability without using folding preferences, as the pretrained model learned this information during its pretraining stage. But I do not feel this would be an important challenge that warrants particular attention in the area. Beyond that, folding information is always better to consider, especially when analyzing protein function.

DPO+KL does not consider the multi-objective setting studied in this work. So, I think the comparison with it is still necessary, as the proposed work claims to use the pretrained IF model as a reference prior. Using the pretrained loss as the regularization for MoDPO objective, as indicated in [3], is another meaningful baseline.

AuthorsOpenReview ↗

Response to Reviewer Xx34

Thanks for your response.

I do not feel this would be an important challenge that warrants particular attention in the area

Answer: We would like to clarify that the practically important problem is how to improve developability through preference alignment without compromising the designed sequences’ ability to fold into the intended backbone. Better solubility or thermostability alone is insufficient if alignment damages structural compatibility. Our experiments directly demonstrate that this trade-off is real and pronounced. Fig. 5(A) shows that property-only DPO progressively degrades TM score, pLDDT, and AAR while improving solubility. By epoch 2, its solubility already exceeds Euterpe’s final value, yet its preservation metrics remain worse. These results establish that pretrained designability is not automatically preserved during property alignment, making its preservation a concrete and practically consequential problem rather than merely an intuitive concern.

folding information is always better to consider, especially when analyzing protein function

Answer: We agree that folding information is important, and we have not discarded it. We use folding-derived metrics to evaluate whether the designed sequences remain compatible with their intended backbones. Our finding is that, without using explicit folding-derived feedback during training, Euterpe achieves comparable folding-based performance to methods that do use such feedback, with better results on some metrics. This does not conflict with the importance of folding. Rather, it shows that its importance for final validation does not make it indispensable as an additional training signal.

Moreover, directly optimizing folding-predictor scores introduces a potential risk of exploiting predictor-specific biases rather than improving genuine structural compatibility. Euterpe does not directly optimize these scores, avoiding that particular route to reward hacking, yet still achieves comparable results under folding-based evaluation. This does not establish that our method is free from all predictor biases, but it strengthens the motivation for separating training supervision from validation. The relevant question is therefore whether omitting folding feedback during training compromises the final designs. Our reported results show that comparable performance can be retained in the evaluated settings, so the general importance of folding alone does not substantiate this concern.

DPO+KL does not consider the multi-objective setting studied in this work

Answer: We would like to clarify that multi-objective design requirements are not equivalent to the number of explicit preference objectives in the loss. For MoMPNN’s TM+Sol setting, our method likewise aims to improve solubility while preserving designability. The difference is that reference-guided preservation replaces explicit TM supervision, so only Sol preferences are required. This does not neglect the structural requirement; it is precisely what our method is intended to achieve. Therefore, DPO+KL should not be considered irrelevant to this setting merely because Sol is its only explicit preference objective. It directly tests whether conventional reference regularization can satisfy the same performance requirements without TM supervision, and Fig. 4 evaluates this alternative across different KL strengths.

Our multi-objective evaluation also extends beyond this setting. Table 1 compares Euterpe [Sol+EP] with MoMPNN [Sol+EP+TM], and Euterpe [Thermo+EP] with MoMPNN [Thermo+EP+TM]. These results demonstrate that our preservation mechanism can be used alongside multiple developability preferences without reintroducing TM supervision. We acknowledge that additional KL- or pretrained-loss-regularized baselines in these more complex settings would further assess the advantages over conventional regularization. However, these would extend the scope of the existing comparisons and should not be conflated with whether the paper addresses multi-objective design or whether the existing DPO+KL comparison is directly relevant.

Comments on the paper

1 comment
Area_Chair_Xsk3OpenReview ↗

Please respond to the rebuttals

Hi reviewers,

Thank you for your reviewing service. The authors have provided fairly extensive rebuttals. Please confirm that you have read them, state which of your concerns are or are not addressed, and whether that changes your recommendations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.