acceptodds
Under review as a conference paper at ICLR 2027

Factorizing LLM-Guided Search for Black-Box Protein Sequence Optimization

Abstract

Protein sequence optimization must search combinatorial landscapes while spending only a small number of costly evaluations. LLMs provide adaptable biochemical priors, but mutation generation alone does not specify how a limited evaluation budget should be allocated across alternative search directions; unconstrained mutation proposals can therefore concentrate on a narrow set of positions and residues. We introduce Factorized LLM Search (FLS), an online optimization method that factorizes search control into parent selection, edit topology, source component, and concrete mutation generation. For each edit topology, FLS updates a two-stage Gibbs policy online, favoring high-reward parents and source components with fewer previously evaluated, chemistry-matched edit directions. A frozen LLM realizes concrete mutations within the sampled search subspaces. We compare FLS with matched prompt-only search, causal controls, and protein-optimization baselines on avGFP, Ube4b, GB1, and Bgl3 using 500 reward-model queries and ten runs per method. FLS improves mean endpoint reward over prompt-only search on all four benchmarks and ranks first in mean endpoint reward on three of four. These results support factorized search as a simple, auditable interface between LLM priors and budgeted protein design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein sequence optimization, large language models, black-box optimization, sequential experimental design

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_cvy3OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes Factorized LLM Search (FLS), a blackbox framework for protein sequence optimization that explicitly separates evaluation allocation from mutation generation. An explicit controller allocates evaluation resources across promising measured variants and feasible local edit components, while a frozen LLM proposes specific amino-acid substitutions. Experiments show the effectiveness of FLS compared with related baselines. Additional ablations and trajectory analyses provide some insights into the respective roles of LLM-based mutation generation and explicit evaluation allocation.

Critical strengths
  • This work is well-motivated and clearly-presented. The paper identifies the conflation of evaluation allocation and mutation generation as a potential limitation of direct LLM-based optimization. Separating these decisions provides a reasonable framework for enhancing the LLM-guided search.

  • The proposed FLS has good interpretability. By introducing an explicit controller while keeping the LLM frozen, FLS makes the allocation process more transparent and facilitates the analysis of search behavior.

  • The empirical evaluation verifies the effectiveness of FLS in protein sequence optimization.

Critical weaknesses
  • The novelty and main insight are not clearly presented. The paper does not sufficiently clarify what new algorithmic mechanisms or theoretical insights the proposed factorization introduces compared with conventional evolutionary search frameworks. The authors should more explicitly identify the fundamental differences and explain how the factorization goes beyond merely reorganizing existing components.

  • The evaluation and analysis are not sufficient. The paper lacks comprehensive evaluations of the individual controller components.

Questions for authors

Please refer to the weaknesses pointed above.

2 comments
AuthorsOpenReview ↗

Response to Reviewer cvy3 [1/2]

Thank you for prompting a more explicit explanation of the factorization insight and the contributions of its components. Following your comments, we provide:

  • Positional analyses to explain the insight behind factorization and distinguish FLS from conventional evolutionary search (W1).
  • Comprehensive ablations to evaluate individual component contributions (W2).

W1. What new algorithmic mechanisms or insights distinguish FLS from conventional evolutionary search beyond reorganizing existing components?

Thank you for your suggestion to clarify FLS’s contribution relative to conventional evolutionary search. Our contribution is to identify a mismatch between fitness variation across position sets and LLM search behavior, and to demonstrate the benefit of separating edit-task allocation from candidate generation. Below, we explain this insight and its implications for positional exploration and endpoint fitness.

Fitness variation across position sets motivates separate allocation

Our central insight is a mismatch between fitness variation across mutation-position sets and the search-budget allocation implicit in LLM proposals. When the LLM jointly proposes mutation positions and residue identities, the allocation of evaluations across positions is implicit in its proposal distribution.

To quantify fitness variation across position sets, we group measured variants by their exact sets of positions mutated relative to the wild type, regardless of residue identities, and retain sets evaluated at least twice. The share of total fitness variance explained by differences between group mean fitness scores is 54–83%. We also measure position-set reuse: the fraction of evaluated candidates whose exact mutated-position set appeared earlier in the same run. For Direct LLM Sampling, 70–91% of candidates reuse such sets while varying residue identities. Both ranges report the minimum and maximum across the four tasks.

To reduce this mismatch, FLS assigns edit tasks programmatically before LLM candidate generation. The analyses below show that FLS explores more positions and achieves higher mean endpoint fitness scores than Direct LLM Sampling.

Relation to conventional evolutionary search. Fitness-based selection determines which individuals survive. In LLM-guided evolutionary methods such as the evaluated MOLLEO baseline, the LLM still chooses where to edit when proposing offspring. FLS makes within-parent task allocation explicit before asking the LLM for a concrete mutation.

Evaluation metrics.

  • Positions visited: distinct positions that differ from the wild type in at least one of a run's 500 evaluated sequences; counts are shown relative to the protein length.
  • Runs covering reference positions: the number of runs whose visited positions include every position at which the benchmark's fixed greedy-search reference sequence differs from the wild type. Visiting all these positions is a necessary condition for generating that reference sequence.
  • Best fitness score: the best fitness score reached within 500 evaluations.

Results. FLS achieves broader positional exploration and higher mean endpoint fitness scores than Direct LLM Sampling on all four tasks. Table W1-1 reports Direct LLM Sampling →\rightarrow FLS after 500 evaluations. Positions visited and reference-position coverage summarize the trajectory analysis over ten runs per method. Endpoint fitness values reproduce the original paper's comparison as mean ± sample SD.

Table W1-1. Position exploration and fitness.

TaskPositions visitedRuns covering reference positionsBest fitness score
avGFP26.7 →\rightarrow 81.7 of 2380/10 →\rightarrow 0/103.966 ± 0.017 →\rightarrow 3.990 ± 0.013
Ube4b32.7 →\rightarrow 69.4 of 1020/10 →\rightarrow 6/105.252 ± 0.879 →\rightarrow 6.938 ± 0.400
GB122.4 →\rightarrow 37.9 of 561/10 →\rightarrow 10/102.835 ± 0.439 →\rightarrow 3.471 ± 0.000
Bgl318.8 →\rightarrow 122.0 of 5010/10 →\rightarrow 2/101.856 ± 0.286 →\rightarrow 2.110 ± 0.117

Together, the analyses support factorized search as a way to reduce this mismatch: FLS visits more positions and achieves higher mean endpoint fitness scores on all four evaluated tasks.

AuthorsOpenReview ↗

Response to Reviewer cvy3 [2/2]

W2. How much do the individual controller and search components contribute to FLS's performance?

To assess the full set of design choices, we compare complete FLS with variants of its controller and candidate-generation components under matched initialization and evaluation budgets. FLS achieves higher mean endpoint fitness scores on at least three of the four tasks in each component comparison. The following controls specify what changes in each variant:

  • Position control: let the LLM choose the source while the program retains parent, edit-action, and chemistry-class control.
  • LLM proposals: replace local LLM proposals with random legal edits within the assigned tasks.
  • Parent selection: replace fitness-based selection with uniform selection from measured variants.
  • Chemistry scheduling: let the LLM choose the chemistry class instead of following the programmed schedule.
  • Recombination: replace the archive-recombination candidate with an additional local LLM proposal, preserving the evaluation budget.

Table W2-1 reports the best observed fitness score after 500 evaluations, as mean ± sample SD over ten runs per method and task.

Table W2-1. Component comparisons.

MethodavGFPUbe4bGB1Bgl3
Complete FLS3.990 ± 0.0136.938 ± 0.4003.471 ± 0.0002.110 ± 0.117
Without position control3.964 ± 0.0146.436 ± 1.0242.744 ± 0.0251.885 ± 0.127
Random local proposals3.992 ± 0.0165.701 ± 1.0163.258 ± 0.3711.873 ± 0.129
Uniform parent selection3.955 ± 0.0126.427 ± 1.0052.771 ± 0.2452.198 ± 0.180
LLM-chosen chemistry class3.985 ± 0.0146.691 ± 0.4682.814 ± 0.4572.210 ± 0.192
Without recombination3.988 ± 0.0235.460 ± 1.4323.137 ± 0.5431.095 ± 0.465

The component effects vary across landscapes: the largest position-control gain occurs on GB1, while LLM proposals and recombination provide their clearest mean endpoint fitness gains on Ube4b and Bgl3. Recombination leverages the mutation pool built up during FLS search, combining previously measured mutations into new candidates. Its effectiveness highlights the value of the mutation options discovered by the other FLS components. Together with the remaining controller comparisons, these results show contributions from multiple FLS components, with their relative impact depending on the landscape.


Thank you again for your thoughtful comments!

Reviewer_9haiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces Factorized LLM Search (FLS) for protein sequence optimization under a limited fitness-evaluation budget. A rule-based controller selects a measured parent based on fitness and assigns a local edit operation, a feasible source position, and an amino-acid chemistry class, while a frozen LLM proposes concrete edits within these constraints. Experimental results show that FLS-generated mutations improve search over random search, genetic algorithm, surrogate-guided, and LLM-guided approaches.

Critical strengths
  1. The paper addresses the important problem of protein sequence optimization and provides useful analysis by separating search allocation from mutation generation.
Critical weaknesses
  1. FLS is closely related to existing LLM-guided evolutionary algorithms, but the paper does not sufficiently establish the novelty or isolate the benefits of its finer-grained control.
  2. End-to-end experiments use only GPT Luna, leaving generalization across LLM backbones unclear.
  3. The presentation needs substantial improvement. Figure 5’s labels and experimental setup are difficult to follow, while the related work provides limited discussion of LLM-based protein optimization and insufficient comparison with the methods evaluated as baselines.
Questions for authors

See Weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 9hai [1/3]

We appreciate your constructive feedback on the novelty, generality, and presentation of FLS. We summarize our responses below:

  • Allocation analyses and ablations to clarify FLS's novelty and the benefits of finer-grained control (W1).
  • Additional LLM backbone experiments to test whether the insight and benefits extend beyond Luna (W2).
  • Presentation clarifications for Figure 5 and expanded discussion of related methods and evaluated baselines (W3).

W1. Compared with existing LLM-guided evolutionary algorithms, what is novel about FLS, and what benefits can be attributed specifically to its finer-grained control?

Thank you for highlighting the need to make this distinction clearer. To address your concern, we present:

  • analyses showing how FLS reduces the mismatch between fitness variation across position sets and LLM-driven budget allocation;
  • an ablation study assessing the contributions of individual FLS components.

Explicit task allocation broadens positional exploration

Distinction from LLM-guided EA. In LLM-guided evolutionary methods such as the evaluated MOLLEO baseline, fitness-based selection determines which individuals are retained in the population, while the LLM proposes offspring without a program-assigned local edit task. FLS additionally assigns the source component and chemistry class before asking the LLM for a concrete mutation. This makes allocation within a selected parent an explicit decision in the search procedure.

Novelty of FLS. Our contribution is to identify a mismatch between where fitness varies and where an LLM spends its search budget, and to reduce this mismatch through factorized search. When the LLM jointly proposes mutation positions and residue identities, the allocation of evaluations across positions is implicit in its proposal distribution.

To quantify fitness variation across position sets, we group measured variants by their exact sets of positions mutated relative to the wild type, regardless of residue identities, and retain sets evaluated at least twice. The share of total fitness variance explained by differences between group mean fitness scores is 54–83%. We also measure position-set reuse: the fraction of evaluated candidates whose exact mutated-position set appeared earlier in the same run. For Direct LLM Sampling, 70–91% of candidates reuse such sets while varying residue identities. Both ranges report the minimum and maximum across the four tasks.

FLS therefore separates edit-task allocation from concrete mutation generation to reduce the concentration of evaluations on position sets favored by the LLM's joint proposals.

Evaluation metrics.

  • Positions visited: distinct positions that differ from the wild type in at least one of a run's 500 evaluated sequences; counts are shown relative to the protein length.
  • Runs covering reference positions: the number of runs whose visited positions include every position at which the benchmark's fixed greedy-search reference sequence differs from the wild type. Visiting all these positions is a necessary condition for generating that reference sequence.
  • Best fitness score: the best fitness score reached within 500 evaluations.

Results. FLS achieves broader positional exploration and higher mean endpoint fitness scores than Direct LLM Sampling on all four tasks. Table W1-1 reports Direct LLM Sampling →\rightarrow FLS after 500 evaluations. Positions visited and reference-position coverage summarize the trajectory analysis over ten runs per method. Endpoint fitness values reproduce the original paper's comparison as mean ± sample SD.

Table W1-1. Position exploration and fitness.

TaskPositions visitedRuns covering reference positionsBest fitness score
avGFP26.7 →\rightarrow 81.7 of 2380/10 →\rightarrow 0/103.966 ± 0.017 →\rightarrow 3.990 ± 0.013
Ube4b32.7 →\rightarrow 69.4 of 1020/10 →\rightarrow 6/105.252 ± 0.879 →\rightarrow 6.938 ± 0.400
GB122.4 →\rightarrow 37.9 of 561/10 →\rightarrow 10/102.835 ± 0.439 →\rightarrow 3.471 ± 0.000
Bgl318.8 →\rightarrow 122.0 of 5010/10 →\rightarrow 2/101.856 ± 0.286 →\rightarrow 2.110 ± 0.117

The results support separating task allocation from LLM mutation generation to reduce positional concentration. In these experiments, FLS visits more positions and achieves higher mean endpoint fitness scores on all four tasks.

AuthorsOpenReview ↗

Response to Reviewer 9hai [2/3]

Component comparisons support the full FLS design

To examine how individual components contribute to these benefits, we compare complete FLS with the following component variants. For each component variant, FLS achieves higher mean endpoint fitness scores on at least three of the four tasks. The controls examine these choices:

  • Position control: let the LLM choose the source while the program retains parent, edit-action, and chemistry-class control.
  • LLM proposals: replace local LLM proposals with random legal edits within the assigned tasks.
  • Parent selection: replace fitness-based selection with uniform selection from measured variants.
  • Chemistry scheduling: let the LLM choose the chemistry class instead of following the programmed schedule.
  • Recombination: replace the archive-recombination candidate with an additional local LLM proposal, preserving the evaluation budget.

Table W1-2 reports the best observed fitness score after 500 evaluations, as mean ± sample SD over ten runs per method and task.

Table W1-2. Component comparisons.

MethodavGFPUbe4bGB1Bgl3
Complete FLS3.990 ± 0.0136.938 ± 0.4003.471 ± 0.0002.110 ± 0.117
Without position control3.964 ± 0.0146.436 ± 1.0242.744 ± 0.0251.885 ± 0.127
Random local proposals3.992 ± 0.0165.701 ± 1.0163.258 ± 0.3711.873 ± 0.129
Uniform parent selection3.955 ± 0.0126.427 ± 1.0052.771 ± 0.2452.198 ± 0.180
LLM-chosen chemistry class3.985 ± 0.0146.691 ± 0.4682.814 ± 0.4572.210 ± 0.192
Without recombination3.988 ± 0.0235.460 ± 1.4323.137 ± 0.5431.095 ± 0.465

Recombination provides the largest mean endpoint fitness gains on Ube4b and Bgl3 by combining previously measured mutations into new candidates. The position-control comparison retains this capability, yet complete FLS achieves higher mean fitness scores on all four tasks. Together, these results support programmed source allocation as a useful addition to the evolutionary search components.

AuthorsOpenReview ↗

Response to Reviewer 9hai [3/3]

W2. Do end-to-end optimization experiments support the effectiveness of FLS with LLM backbones other than GPT Luna?

To test whether the insight extends across LLM backbones, we applied the same trajectory analyses to DeepSeek V4.1 Flash and Gemini 3.8 Flash. Both backbones reproduce the mismatch and show broader positional exploration with FLS on all four tasks.

Mismatch diagnosis. Differences between mutated-position sets account for 51.8–80.8% of observed fitness variance in the DeepSeek samples and 62.9–78.6% in the Gemini samples. Yet 57.7–87.3% of Direct LLM Sampling candidates with DeepSeek and 63.4–87.6% with Gemini reuse previously explored position sets while varying residue identities. FLS lowers this position-set reuse rate on every task with both backbones.

Evaluation setup. We follow the evaluation protocol and trajectory metrics in W1, using the first five seeds for each additional backbone, matched initializations across FLS and Direct LLM Sampling, and 500 fitness evaluations per run.

Results. Tables W2-1 and W2-2 report Direct LLM Sampling →\rightarrow FLS for DeepSeek and Gemini, respectively, using the reporting conventions in W1. FLS visits more positions on all four tasks with both backbones, and also achieves higher mean endpoint fitness scores on Ube4b and GB1 with both workers and on Bgl3 with Gemini.

Table W2-1. DeepSeek exploration and fitness.

TaskPositions visitedRuns covering reference positionsBest fitness score
avGFP30.4 →\rightarrow 84.0 of 2380/5 →\rightarrow 0/54.002 ± 0.014 →\rightarrow 3.981 ± 0.012
Ube4b44.4 →\rightarrow 80.8 of 1020/5 →\rightarrow 5/55.932 ± 1.102 →\rightarrow 6.475 ± 1.703
GB136.2 →\rightarrow 49.8 of 561/5 →\rightarrow 5/52.562 ± 0.000 →\rightarrow 3.218 ± 0.230
Bgl351.4 →\rightarrow 138.4 of 5010/5 →\rightarrow 0/51.898 ± 0.144 →\rightarrow 1.896 ± 0.114

Table W2-2. Gemini exploration and fitness.

TaskPositions visitedRuns covering reference positionsBest fitness score
avGFP25.8 →\rightarrow 55.6 of 2380/5 →\rightarrow 0/54.000 ± 0.005 →\rightarrow 3.994 ± 0.008
Ube4b62.4 →\rightarrow 63.0 of 1021/5 →\rightarrow 3/55.153 ± 2.159 →\rightarrow 6.421 ± 0.670
GB132.8 →\rightarrow 37.8 of 565/5 →\rightarrow 5/52.744 ± 0.406 →\rightarrow 3.471 ± 0.000
Bgl336.8 →\rightarrow 110.6 of 5011/5 →\rightarrow 0/51.992 ± 0.312 →\rightarrow 2.134 ± 0.013

Additional trajectory diagnostic. New-position rate is the fraction of queries 2–500 that introduce at least one position absent from all preceding evaluated sequences. Rates are averaged across runs for each task. Across tasks, rates increase from 2.8–10.0% to 8.8–27.5% with DeepSeek and from 1.9–9.9% to 6.4–21.9% with Gemini.

Together, these findings extend the W1 insight beyond Luna and support separating edit-task allocation from LLM mutation generation to reduce the mismatch in search-budget allocation.

W3. How will Figure 5's presentation and the discussion of prior LLM-based protein optimization methods and evaluated baselines be improved?

We sincerely apologize for the lack of clarity in the presentation. We will revise Figure 5 and its caption to clarify the labels and experimental setup, and make the main finding explicit in the caption.

Revised caption (excerpt). Figure 5 shows that LLMs can recover high-scoring reference variants under specified edit constraints. However, even from a starting variant only one edit operation away from the reference, they do not reliably recover it with a single proposal, particularly when relocation is required.

We have also expanded the discussion of LLM-based protein optimization and explicitly compared FLS with the methods evaluated as baselines. The revised discussion distinguishes protein optimization methods from those developed for molecular design and other domains, and explains how their search mechanisms differ from FLS's programmed edit-task allocation.


Thank you again for your careful review and guidance!

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

This paper proposes FLS for black-box protein sequence optimization. The main idea is to separate evaluation allocation from mutation generation: an explicit controller selects promising parent sequences and feasible local edit components, while a frozen LLM proposes concrete amino-acid substitutions. Experiments on four protein fitness landscapes show competitive best-of-budget performance, and controlled ablations investigate the contributions of position control and LLM-based mutation proposals.

Critical strengths

The paper identifies an interesting distinction between mutation generation and search allocation, which provides a useful perspective on LLM-guided optimization.

The proposed framework is intuitive and interpretable, with a clear separation between the search controller and the LLM worker.

The experiments include useful ablation studies and analyses of search trajectories, providing evidence that explicit position control can improve exploration.

The paper is clearly written and generally transparent about its methodological limitations.

Critical weaknesses

The paper emphasizes the importance of deciding where to search, but the proposed strategy is relatively simple. For example, mutation positions are selected uniformly at random. It would be helpful to compare this approach with other search strategies to better understand its advantages.

While the ablations show that controlling mutation positions is useful, it is still unclear how much each part of the method contributes to the final results. In particular, the effects of chemistry scheduling and recombination could be studied further.

Some baselines use different initial sequences, which may affect the results. Using the same initialization across methods would make the comparisons more convincing.

Questions for authors

Please refer to the weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer DEZe [1/2]

Thank you for your careful assessment and practical suggestions for strengthening the experimental comparisons. Following your comments, we provide:

  • Rationale for uniform source sampling and clarification of its role in FLS (W1).
  • Expanded component ablations to quantify individual contributions to final fitness (W2).
  • Additional matched-initialization comparisons to test whether the original conclusions hold under shared starting conditions (W3).

W1: How does FLS's simple uniform position/source selection rule compare with alternative search strategies, and what advantages does it offer?

Thank you for raising this point. We would like to clarify that uniform source sampling is a simple implementation choice rather than a core contribution of FLS. Our contribution is to decouple position allocation and chemistry scheduling from the LLM's concrete mutation proposals, making these allocation decisions explicit and independently controllable.

With a limited evaluation budget, the measured data provide only sparse observations of the fitness landscape, making it difficult to reliably estimate the potential of unmeasured edit opportunities. We therefore use uniform sampling over feasible sources as a simple default that does not require such estimates.

Alternative search strategies could incorporate more informative priors for source allocation, but developing such strategies is beyond the scope of this work.

W2. How much does each part of the method contribute to the final results, particularly chemistry scheduling and recombination?

To quantify how much each component contributes to the final results, we compare every variant directly with complete FLS under matched initialization and evaluation budgets. FLS achieves higher mean endpoint fitness scores on at least three of the four tasks in each component comparison. The controls include the chemistry schedule and recombination you highlighted:

  • Chemistry scheduling: let the LLM choose the chemistry class instead of following the programmed schedule.
  • Recombination: replace the archive-recombination candidate with an additional local LLM proposal, preserving the evaluation budget.
  • Position control: let the LLM choose the source while the program retains parent, edit-action, and chemistry-class control.
  • LLM proposals: replace local LLM proposals with random legal edits within the assigned tasks.
  • Parent selection: replace fitness-based selection with uniform selection from measured variants.

Table W2-1 reports the best observed fitness score after 500 evaluations, as mean ± sample SD over ten runs per method and task.

Table W2-1. Component comparisons.

MethodavGFPUbe4bGB1Bgl3
Complete FLS3.990 ± 0.0136.938 ± 0.4003.471 ± 0.0002.110 ± 0.117
Without position control3.964 ± 0.0146.436 ± 1.0242.744 ± 0.0251.885 ± 0.127
Random local proposals3.992 ± 0.0165.701 ± 1.0163.258 ± 0.3711.873 ± 0.129
Uniform parent selection3.955 ± 0.0126.427 ± 1.0052.771 ± 0.2452.198 ± 0.180
LLM-chosen chemistry class3.985 ± 0.0146.691 ± 0.4682.814 ± 0.4572.210 ± 0.192
Without recombination3.988 ± 0.0235.460 ± 1.4323.137 ± 0.5431.095 ± 0.465

For the two components you mentioned, programmed chemistry scheduling yields higher mean endpoint fitness scores than LLM-chosen classes on three of the four tasks. Recombination provides the largest mean endpoint fitness gains on Ube4b and Bgl3. It leverages the mutation pool built up during FLS search, combining previously measured mutations into new candidates. Its effectiveness highlights the value of the mutation options discovered by the other FLS components.

AuthorsOpenReview ↗

Response to Reviewer DEZe [2/2]

W3. Do comparisons between FLS and the relevant baselines still support the paper's conclusions when all methods use the same initial sequences?

We agree that initialization can influence subsequent proposals and final performance. Below, we would like to:

  • P1. clarify the rationale for the original initialization protocols
  • P2. report additional reruns with shared starting conditions that show the original conclusions are preserved.

Explanation of Different Initialization (P1). FLS is a self-contained search algorithm that can begin optimization without an externally supplied set of measured variants. Given the wild-type sequence and task constraints, it uses the LLM's prior knowledge to generate its first batch of candidates before observing any fitness measurements; this batch therefore reflects the LLM's prior mutation preferences. For baselines that require an external initial pool, the supplied variants and measured scores provide the task-specific evidence needed to initiate their optimization procedures. The original initialization protocols thus reflect a methodological difference in how the methods obtain their starting information.

Shared Initialization Experiments (P2). We agree that a shared initialization provides a more controlled comparison of subsequent search performance. Following your suggestion, we reran two major baselines, MOLLEO and Kermut, using the same initial sequences and measured fitness scores as FLS for each task and run block. The shared-initialization results in Table W3-1 preserve the original comparative pattern.

Experimental Setup. Each rerun starts from the same five measured variants as its corresponding FLS run, with these initial measurements counted inside the total budget of 500 fitness evaluations. Table W3-1 reports the original baseline results alongside the matched-initialization reruns. The FLS row is the original reference whose initialization is shared by the reruns.

Table W3-1. Matched-initialization baseline reruns.

MethodInitializationavGFPUbe4bGB1Bgl3
FLSShared reference3.990 ± 0.0136.938 ± 0.4003.471 ± 0.0002.110 ± 0.117
MOLLEOOriginal baseline starts3.923 ± 0.0174.463 ± 1.6072.930 ± 0.5762.099 ± 0.308
MOLLEOMatched to FLS3.922 ± 0.0115.650 ± 0.9653.143 ± 0.4351.981 ± 0.379
KermutOriginal baseline starts3.947 ± 0.0268.070 ± 0.2233.170 ± 0.4042.008 ± 0.105
KermutMatched to FLS3.987 ± 0.0207.985 ± 0.2053.300 ± 0.3601.990 ± 0.178

Values are the best observed fitness score after 500 evaluations, reported as mean ± sample SD over ten runs per method and task.

The shared-initialization results preserve the original comparative pattern: FLS achieves higher mean endpoint fitness scores than MOLLEO on all four tasks and than Kermut on three of the four tasks. We will include the completed matched-initialization comparison in the revision.


Thank you again for your constructive suggestions!

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to the AC and All Reviewers [1/1]

We thank the AC and the reviewers for their careful and invaluable feedback. We are encouraged that the reviewers recognized the following strengths:

  • This work addresses an important and well-motivated problem, using LLMs to search for improved protein sequences under a limited evaluation budget.
  • Separating search allocation from mutation generation provides a useful perspective on LLM-guided optimization.
  • The explicit controller and frozen LLM provide an intuitive and interpretable framework.
  • The experiments demonstrate the consistent effectiveness of FLS for protein sequence optimization.
  • The useful ablations and trajectory analyses provide insights into the roles of explicit position control and LLM mutation generation.
  • The paper provides methodological transparency in describing the method's scope.

Following the reviewers' suggestions, we have conducted additional experiments and expanded our analyses to address the main concerns:

  • Contribution and novelty. Our analyses identify a mismatch between fitness variation across mutation-position sets and the budget allocation implicit in LLM proposals. FLS separates edit-task allocation from mutation generation to reduce this mismatch, with trajectory analyses showing less position-set reuse and broader positional exploration.
  • Additional Component Ablations. We extend the original analyses with comparisons for chemistry scheduling and recombination. Complete FLS achieves higher mean endpoint fitness scores on at least three of the four tasks in each component comparison.
  • Additional Matched-Initialization Comparisons. We rerun MOLLEO and Kermut using the same initial sequences and measured fitness scores as FLS. The results preserve the original comparative pattern.
  • Different LLM backbones. We repeat the trajectory analyses and optimization comparisons with DeepSeek V4.1 Flash and Gemini 3.8 Flash. FLS reduces position-set reuse and visits more positions on all four tasks with both workers. It retains its fitness advantages on Ube4b and GB1 across both additional backbones, with further gains on Bgl3 using Gemini.

In addition, we will clarify the scope and implementation of FLS in the revision, including the role of uniform source sampling and the rationale for the initialization protocols. We will also revise Figure 5 and its caption, and expand the discussion of related LLM-based protein optimization methods and evaluated baselines.

We hope these additions address the reviewers' concerns, and we are happy to answer any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.