acceptodds
Withdrawn from ICLR 2027

RESOLVE: Jointly Valuing Model Adaptation and Experiments for Protein Engineering

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Protein engineering requires choosing which variants to measure and how to adapt predictive models from limited experimental data. Nonmyopic experimental design values measurements for both immediate discovery and their ability to improve later choices. When models are fine-tuned on new observations, these choices also depend on changing representations. We introduce RESOLVE, a controller that combines acquisition lookahead with a common discovery objective for model adaptation and experimental selection. Its evaluator accounts for representation changes through hypothetical predictor and representation updates and assesses frozen and adapted recommendations through the experimental batches they propose. Our theoretical analysis connects valuation accuracy to immediate selection quality. Retrospective experiments across diverse protein assays and feedback schedules demonstrate improved discovery over acquisition and planning baselines under matched measurement budgets. Ablation studies show that accounting for adapted representations improves value assessment, while selecting between frozen and adapted recommendations improves discovery. These results support representation-aware valuation and coupled action-batch selection as useful components of nonmyopic experimental design for protein engineering.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein engineering, Nonmyopic active search, Sequential experimental design, Representation adaptation, Active learning, Value of information

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_U9NJOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

In general, this paper wants to solve the problem of sequential selection among a finite set of protein variants, aiming to identify more meaningful variants within a limited budget. The authors propose a method called RESOLVE, which performs different selection steps using a shared backbone. They validate the effectiveness of the proposed method across several tasks and show some good results; I think the experimental evaluation is comprehensive.

Critical strengths

The experimental evaluation is comprehensive, covering multiple tasks and settings; the proposed method demonstrates effectiveness, yielding improvements and gains over the baseline approaches.

Furthermore, the research problem of this paper is clearly defined, and variables were well-controlled.

Critical weaknesses

The proposed method may be overly complex and lacks comparisons with simple baselines (or compute-matched baseline); the authors should analyze and justify whether such a complex control system is truly necessary (or at least give some discussion).

Furthermore, I suggest that the authors provide a detailed discussion and analysis regarding the various components and the selection of hyperparameters (and the reasons), which might benefit their work.

Questions for authors

See Critical Weaknesses and,

  1. Can the authors report the actual costs—specifically, whether the performance gains of the proposed method come from increased computational complexity and higher costs?

  2. Is there a compute-matched baseline?

  3. Could you further elaborate more on the independent contribution of the controller?

Reviewer_P4K3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how to select among experimental batches proposed by frozen and fine-tuned PLM. RESOLVE employs a Gaussian process ensemble and simulated model updates to evaluate the discovery value of each batch, including its impact on future experiments. Results across 24 ProteinGym experiments show that this method outperforms fixed-perspective strategies and various acquisition baselines in elite recall rate.

Critical strengths

1.This paper addresses an important issue in protein engineering: under the condition of limited experimental budget, when is it worthwhile to use model adaptation. Compared with the improvements achieved by always freezing and always adjusting strategies, it indicates that making a choice between the two has practical value.

  1. The evaluation is comprehensive, covering 24 experiments and six feedback plans. Comparisons with self-scoring, head repetition, and other internal variants help assess the effectiveness of the proposed components.
Critical weaknesses
  1. The external comparison was not fully controlled. Multiple benchmark methods used different predictors and training objectives, while RESOLVE required repeated Gaussian process fitting and LoRA updates. Without reporting the computational cost, it is difficult to determine whether this improvement is sufficient to justify the additional expense.

  2. The contribution of the proposed forward mechanism is still unclear. In the shared state analysis, its advantage over Own-Score was only 0.06 hits, and the confidence interval crossed zero. Additionally, the ablation experiments on representation updates also changed the feature coverage range, making it difficult to quantify its individual contribution.

Questions for authors

1.Could the authors report the computational cost per campaign and compare against baselines under a similar computational budget?

2.Can the effect of representation updates be separated from changes in feature coverage? Why is the gain over Head-Repeat only 0.17 percentage points on the eight shortlist assays?

3.Section 4.1 describes ten epochs of LoRA training, while Appendix A.4 specifies ten optimizer steps. Which setting was actually used?

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

This submission introduces RESOLVE, a framework for iterative protein optimization that decides between leveraging representations of either frozen or fine-tuned pLMs based on immediate and simulated performance in subsequent rounds. The authors evaluate their method on retrospective assays from ProteinGym.

Critical strengths

The authors include code from their experiments, which is commendable

Critical weaknesses

W1. Clarity and presentation quality.

Central weakness of this submission is that it is very poorly written and presented, to the extent that it feels mostly AI generated. Terminology specific to the framework is notoriously introduced without prior definition or explanation, subsequent paragraphs lack consistency and/or logical structure, and the figures are cluttered with AI-style captions and annotations that make it extremely difficult to understand what they are intended to convey. In my view, the severity of this issues undermine accessibility of the submission, and consequently, its relevance to the community.

W2. Lack of proper evaluation under covariate shifts

Particularly for a lookahead-based acquisition strategies, performance of any iterative protein optimization framework depends strongly on the quality of its surrogate predictors. In practice, these models tend to be misspecified due to the covariate shift between observed and candidate sequences. Because of this, majority of the experiments conducted in this work don't convince me that the reported performance improvements would generalize to other settings, since currently the evaluation is restricted to fitness landscapes with fixed mutational sites, mostly single-mutant/mixed at that.

Questions for authors

Q1. View composition

Why did the authors construct each "view" such that it consists of those specific 10 GPs? How does the performance depends on the number of different "worlds", and different kernels?

Q2. Compute comparison

What's the total compute cost of your method in comparison to other methods, in particular other non-myopic ones?

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

Related papers

Discussion (0)

Sign in to comment.