acceptodds
Under review as a conference paper at ICLR 2027

EvoCircuit: Active Learning for Protein Optimization via Mechanistic Interpretability

Abstract

Protein language models (pLMs) provide rich representations for protein engineering, but active optimization must adapt these representations to target-specific objectives from only a small number of experimental measurements. Existing protein optimization approaches improve candidate selection through activity prediction or search over learned fitness landscapes, but do not explicitly exploit task-relevant sparse representations within the underlying pLM. We introduce EvoCircuit, an active-learning framework informed by mechanistic interpretability that uses sparse features extracted from cross-layer transcoders of ESM-2. At each round, EvoCircuit identifies task-relevant features through associations between feature activations and experimental activity in measured variants. These features induce a task-adaptive similarity over candidate sequences, guiding batch acquisition alongside predictor-based ranking. Across retrospective active-learning simulations on single-mutant protein fitness landscapes, EvoCircuit improves the discovery of high-activity variants over predictor-based selection under matched assay budgets, with effects varying across landscapes and predictors. Ablations examine the contributions of task-adaptive feature selection and acquisition design, including comparisons with embedding-based and random-feature controls. Experiments on a combinatorial multi-mutant landscape extend the evaluation to predictor-based and reward-guided search baselines, revealing differences between hit discovery and peak-fitness optimization. Our results suggest that sparse internal representations of pretrained pLMs can complement activity prediction as inductive signals for data-efficient protein optimization.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
mechanistic interpretability, applications to biology, Protein Language Models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_AseaOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This submission introduces EvoCircuit, a framework for protein optimization where the candidate selection is additionally guided by the proximity to sparse features of high-scoring experimentally validated sequences extracted from cross-layer transcoders of ESM-2. The authors evaluate their method on 7 protein fitness landscapes through retrospective active-learning simulations.

Critical strengths

The idea to incorporate sparse features extracted from pretrained pLMs into protein optimization is novel, and the paper is nicely written and easy to follow.

Critical weaknesses

The paper has following weakness:

W1 Contribution is limited

I do not understand why the authors split their contributions the way they do, namely into (i) the introduction of EvoCircuit, (ii) its evaluation, and (iii) ablation studies. I would expect that introducing a framework naturally entails evaluating it and conducting ablations; the latter two do not seem to constitute standalone contributions. This gives the impression that the main contribution is a somewhat hand-wavy implementation of the idea of incorporating sparse features into candidate selection, which, in my view, is rather incremental. In such cases, I would be more convinced by an extensive experimental evaluation demonstrating strong performance against a diverse set of baselines. However, I find the baseline comparisons limited and the gains achieved by EvoCircuit modest.

W2 Evaluation is weak and limited to fitness landscapes with fixed mutational sites

Having said that the idea of incorporating sparse features extracted from pretrained pLMs into protein optimization is novel, I am not convinced it's a good and effective one, and experiments on fitness landscapes in the paper do little to change my mind. Given that the proposed method uses sparse features to guide candidate retrieval rather than relying solely on surrogate predictions, it would be beneficial to demonstrate gains on fitness landscapes which don't cover 90%+ of the variant space, where surrogate models tend to be misspecified and having such additional signal would be extremely valuable. In the experiments conducted by the authors, I remain unconvinced that the proposed approach provides substantial benefits, particularly given the modest gains achieved by EvoCircuit.

W3 Authors did not provide code, and there is no reproducibility statement

Questions for authors

Q1: Could the authors conduct some experiments on fitness landscapes without fixed mutational sites? See for example landscapes used in [1-3]

Q2: What's the diversity of sequences returned by your method, and how does it change given the number of different anchors?

Q3: How does the performance depends on the number of selected features?

I understand this requests a lot of experiments, but given the limited in my eyes contribution of this submission, I believe this is the only angle which could make this work potentially relevant for the community. In particular, I will consider raising my score if Q1 will be adequately addressed.

References

[1] Kirjner et al., Improving protein optimization with smoothed fitness landscapes, ICLR 2024

[2] Kmicikiewicz et al., ProSpero: Active Learning for Robust Protein Design Beyond Wild-Type Neighborhoods, NeurIPS 2025

[3] Ren et al., Proximal Exploration for Model-guided Protein Sequence Design, ICML 2022

4 comments
AuthorsOpenReview ↗

Expanded Evaluation Beyond Fixed-Site Benchmarks

Thank you for your careful review and constructive suggestions.

[W1 — Clarifying the contribution]

We revised the contribution statement in the Introduction to emphasize the acquisition framework and empirical findings, rather than evaluation and ablation as standalone contributions. It now highlights predictor- and landscape-dependent discovery gains across eleven landscapes (Sections 5.1–5.2) and feature-level analysis linking activation patterns to recommendations through feature removal (Section 5.4). Section 5.3 separately provides matched fixed-state evidence for the contribution of activity-based feature selection.

[W2 / Q1 — Evaluation beyond fixed-site benchmarks]

Following your suggestion and the references you shared, we added Pab1, GFP, cgreGFP, and AAV to evaluate libraries with variable mutation-site subsets (Sections 4.1 and 5.2, Table 4). Pab1 contains single and double mutants; the other three also include higher-order variants. GFP and cgreGFP vary mutation sites across their sequences, while AAV varies subsets within a 28-residue region.

Each comparison uses the same predictor configuration, ten paired seeds, 64 shared uniform-random initial measurements, and 112 additional variant selections. Across the full measured pools, EvoCircuit increases mean top-10% discoveries in 17 of 20 comparisons against greedy acquisition. On cgreGFP/FolDE, discoveries increase from 20.5 to 29.9 (+45.9%), and on AAV/XGB from 48.2 to 55.7 (+15.6%).

Unlike its mixed fixed-site results, FolDE with EvoCircuit improves all three discovery metrics—top-10% count, top-1% count, and discovery AUC—on GFP, cgreGFP, and AAV. For DNN and XGB, EvoCircuit likewise increases the mean values of all three discovery metrics on these libraries. While gains remain setting-dependent, these results show discovery benefits beyond near-complete combinations of a few fixed sites.

[Q2 — Discovery coverage and anchor count]

Appendix J measures distinct-hit coverage across repeated campaigns rather than within-campaign sequence distance. On TrpB4/RF, the union of distinct above-parent variants across 20 campaigns increases from 302 to 352 (+16.6%), with nearly unchanged mean per-campaign discoveries (greedy: 64.6; EvoCircuit: 64.4). Across the four added libraries, the 10-campaign union of top-10% variants increases in 16 of 20 matched comparisons. These results show broader discovery coverage, including less repeated discovery in the TrpB4/RF setting.

We also tested A = 4, 16, and 32 anchors with RF on GB1 and TrpB4, fixing K = 8, shortlist size M = 128, and the direct predictor/feature-guided allocation at 8+8. Each setting used 10 paired seeds, the parent plus 63 shared random initial variants, and 112 additional variant selections. Increasing A distributed the fixed shortlist budget across more high-activity neighborhoods, without producing a clear increase in sequence diversity (normalized Hamming distance) in these tests. These anchor- and feature-count analyses are reported here rather than in the current revision.

[Q3 — Sensitivity to the number of selected features]

In the same RF study, we varied the maximum number of selected features K while keeping A = 16 anchors. Mean above-parent discovery AUCs are:

DatasetK = 4K = 8K = 32
GB134.733.436.7
TrpB425.926.226.8

An exploratory DNN/one-hot follow-up on GFP and AAV used 10 paired seeds, 64 shared uniform-random initial measurements, and 112 additional variant selections. Increasing K from 8 to 32 raises top-10% discovery AUC from 17.6 to 21.5 on AAV (paired gain 3.9; pointwise 95% CI [1.9, 6.1]) and from 10.3 to 12.6 on GFP (paired-difference 95% CI [-0.1, 4.8]). Among the sensitivity contrasts reported here, only the AAV/DNN increase is statistically significant after Holm correction within the respective four-contrast families.

We retain K = 8 as a common reference, not a claimed optimum. The response to K varies by setting, and the AAV/DNN result shows that a larger feature set can further improve acquisition.

[W3 — Implementation and reproducibility]

We plan to release the implementation and feature-exploration demo with the camera-ready version. Appendices A, B, and D specify the architecture, acquisition settings, and predictor configurations; Appendix L illustrates our demo. These details and the planned code release are intended to support independent reproduction of the experiments and inspection of candidate-selection decisions.

Taken together, the expanded evaluation shows that EvoCircuit can complement existing predictors in discovering high-activity variants beyond the original fixed-site benchmarks. We welcome further discussion of these results and their interpretation.

Thank you for providing this additional analyses. The variable-site experiments are a step in the right direction, and do strengthen the paper. However, my previous concerns about value added by EvoCircuit remain not fully addressed.

On contribution

My original comment was meant to ask for a clearer demonstration of what EvoCircuit contributes and why that contribution matters. The revised introduction does describe the contribution more precisely in my view, mainly through empirical performance of the proposed method. As such, the strength of EvoCircuit performance remains central to my assessment.

On performance

In my view, the new experiments do not establish how competetive/usefull EvoCircuit is against more commonly used for variable-site protein search methods. Greedy acqusition is a useful control, but I give very little weight to gains over TS. With sparse coverage on those types of landscapes combined with small measurement budget, TS cannot reliably guide selection because it's stuck in its exploratory stage, and cannot start exploiting. In my eyes the paper still lacks a convincing comparison against more established methods used on variable-site landscapes (say, PEX, AdaLead, GFNs, etc), or demonstration of a setting where the value added by EvoCircuit is crucial for the performance (say, distribution shift or some other setting), where other methods fail. Additionally, here the gains are even harder to assess, because Table 4 (and 3 also) lack standards deviations / confidence intervals

Taken together, I appreciate the effort made by the authors, however given the current evidence I will be keeping my current score of weak rejection

AuthorsOpenReview ↗

Analysis of EvoCircuit’s Added Value

Thank you for encouraging stronger comparisons with established methods on variable-site landscapes and a clearer analysis of EvoCircuit’s added value.

Overall, EvoCircuit can complement existing acquisition policies with activity-associated CLT features. Our additional experiments identify both the candidates this signal recovers and the features influencing their selection. We also report SDs and 95% CIs for each method’s mean in Tables 3 and 4 (Appendix N).

[Comparison with established methods]

We compared and combined EvoCircuit with PEX, which you suggested as an established method for variable-site landscapes (Tables L.1–L.2). Initial observations and assay budgets were matched across four libraries and six predictors. Adding EvoCircuit increases mean top-1% discoveries in 21/24 combinations and discovery AUC in 15/24. On GFP, TabPFN—the strongest unaugmented PEX configuration by mean AUC—improves from 5.2±2.9 to 7.0±3.1 top-1% discoveries (paired-difference 95% CI [0.5, 3.1]).

[What CLT-guided selection contributes]

EvoCircuit supplies additional selection evidence for high-activity candidates insufficiently prioritized by the existing predictor and policy. With observed measurements and predictor scores held fixed, we compare the numbers of top-10% variants selected exclusively with EvoCircuit and exclusively by the baseline policy.

EVOLVEpro acknowledges that greedy selection may overlook fitness-landscape regions [1]; our RF/ESM-2 Pab1 comparison exhibits positive net recovery beyond its top-ranked candidates, alongside campaign discoveries increasing from 44.2 to 48.5. FolDE reports that top-N batches concentrate on previously successful mutation sites [2]. With our FolDE/ESM-2 adaptation, CLT retrieval recovers high-activity candidates outside the top-ranked set, providing another selection basis even when the predictor already uses pLM information. ALDE finds that global prediction accuracy and calibration need not track optimization success [3]. Our DNN/XGB results likewise identify benefit through actual candidate recovery: with Greedy and the FolDE/DNN/XGB-derived predictors, low-order variants (≤2 mutations) account for most additional top-10% discoveries in all nine GFP/cgreGFP/AAV combinations.

PEX explicitly prioritizes high-fitness candidates with fewer mutations [4]. Adding EvoCircuit to PEX/RF increases mean higher-order top-10% discoveries (≥3 mutations) across GFP, cgreGFP, and AAV. These results support EvoCircuit as an orthogonal acquisition approach that can yield additional performance gains when combined with existing predictors and search policies (Figures M.1–M.2).

[Interpreting acquisition decisions and their significance]

These selections can be traced to the features used by acquisition. In a GFP example, an activity-associated feature admits a top-1% candidate omitted by PEX; removing that feature from retrieval excludes the candidate (Appendix L). Following InterPLM’s interpretability workflow [5], we examine feature activations in relation to biological annotations. In GCK, we inspect local correspondence with the annotated substrate-binding residue N204 and separately verify the feature’s influence on candidate selection (Appendix J, Tables J.1–J.2).

Although this does not establish a biochemical mechanism, this connection makes the feature evidence behind acquisition decisions inspectable and provides a basis for formulating concrete hypotheses for follow-up experiments under limited assay budgets.

EvoCircuit therefore offers a framework that connects improved candidate discovery with feature-level understanding of acquisition decisions in protein optimization.


[References]

[1] Jiang et al., Rapid in silico directed evolution by a protein language model with EVOLVEpro, Science, 2025.

[2] Roberts et al., Low-N Protein Activity Optimization with FolDE, arXiv:2510.24053, 2025.

[3] Yang et al., Active learning-assisted directed evolution, Nature Communications, 2025.

[4] Ren et al., Proximal Exploration for Model-guided Protein Sequence Design, ICML, 2022.

[5] Simon and Zou, InterPLM: discovering interpretable features in protein language models via sparse autoencoders, Nature Methods, 2025.

Thank you for the additional experiments and the clarification. Revised Appendix J is much more interesting. Regardless of the outcome here, I think demonstrations and connections like this are directions that could make this work potentially exciting for the community. I will change my score to 3.

Reviewer_8UkLOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes EvoCircuit, a method for choosing which protein variants to test next in a lab campaign. It finds a few internal features of a protein language model (ESM-2) whose values track the activity measured so far. It then tests unmeasured variants that look like the best hits along those features. Experiments are conducted on different measured protein fitness landscapes, and the gain is often modest.

Critical strengths

The science problem the paper studies is important for saving laboratory efforts. The paper seems to be studying the latest pLMs model (according to related work) and is making progress on it. To me, the paper explained the general idea clearly, and experiments over different settings are conducted.

Critical weaknesses

I am not an expert in this area and am unable to assess the correctness of the scientific claims within the paper. To me, the method is generally reasonable, but I have concerns especially on the experiment part of the paper.

  1. Table 1 only shows small and noisy gains of EvoCircuit. The per-dataset standard deviations are very large, and the benefit of EvoCircuit is relatively small. Though averaged over five seeds, this instability limits the practical usage of the method. Also, under FolDE, the strongest predictor, EvoCircuit is essentially a tie. This raises questions on whether the proposed method is really needed, since a stronger predictor would always be more preferred.
  2. Section 5.4 is written at a very low level of detail, naming individual residues, variants and rankings. This may suit a specialist audience, but the section should state up front what the case study is meant to show and what the authors take away from it, rather than leaving readers to do the work.
Questions for authors
  1. The paper misses to cite necessary things, such as XGBoost.
  2. In Appendix section K.1, algorithm and equation references are missing.
1 comment
AuthorsOpenReview ↗

Feature-Guided Acquisition and Interpretability

Thank you for your thoughtful review and helpful feedback.

[W1 — Discovery gains and baseline comparisons]

In the single-mutant experiments, gains depend on the predictor and landscape; FolDE achieves higher mean discovery counts than the RF/TabPFN configurations in Table 2. EvoCircuit is a complementary acquisition strategy that can be integrated with existing predictors without changing their architectures. Its contribution is therefore distinct from improving the predictor itself: it uses activity-associated CLT features to guide which candidates are evaluated next. To assess this added value, we extended the evaluation to Pab1, GFP, cgreGFP, and AAV (Section 5.2, Table 4), using ten paired seeds per library, 64 shared uniform-random initial measurements, and 112 additional variant selections.

Across these libraries, EvoCircuit increases mean top-10% discoveries in 17 of 20 comparisons against greedy acquisition with the same predictor configuration. On cgreGFP, FolDE increases discoveries from 20.5 to 29.9 (+45.9%) with EvoCircuit, exceeding both greedy selection and its constant-liar acquisition rule (22.4). These comparisons use the matched ESM-2 adaptation described in Appendix D. On AAV, XGB achieves 55.7 discoveries with EvoCircuit, compared with 48.2 under greedy selection and 37.5 under Thompson sampling.

While discovery gains do not consistently translate into higher peak fitness (see Section 6), these comparisons show that EvoCircuit can improve high-activity variant discovery without changing the predictor configuration, including relative to non-greedy acquisition policies.

[W2 — Clarity and practical value of feature interpretation]

We revised Section 3 and Figure 1 to clarify the acquisition procedure, and Section 5.4 to make the case study accessible without detailed biological background. The same activity-associated features used to guide acquisition can also be examined in biological context through reference-protein descriptions and residue-level activation profiles.

In the GCK example, the feature shows strong activation at glutamate 221, consistent with the residue preference described from reference proteins. This feature also guided the retrieval of D217A, a top-10% variant not selected by predictor-only ranking. Holding measurements and predictor scores fixed, removing it excludes D217A and reduces top-10% hits among the eight feature-guided selections from three to one. The eight variants selected through direct predictor ranking remain unchanged.

Together, these analyses connect the biological interpretation of a feature to its role in selecting a high-activity variant. While they do not establish the biochemical mechanism underlying that variant's activity, they provide a basis for biological follow-up on features actually used in acquisition. Detailed activation profiles and replay results are provided in Appendix J, with feature-removal results in Table J.2.

[Q1 / Q2 — Citation and algorithm references]

We added the missing XGBoost citation (Chen and Guestrin, 2016). The replay in Appendix J.1 follows the acquisition rule in Section 3.3: Eq. (3) defines the candidate–anchor distance, and Eq. (4) defines the selection for each round.

Overall, the expanded comparisons support the discovery value of feature-guided acquisition with existing predictors, while the case study demonstrates how individual variant selections can be examined through the features used in retrieval. We welcome further questions or requests for clarification.

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

The paper proposes an approach that leverages the analysis of different activation values coming from mechanistic interpretability frameworks to create associations between sparse features extracted from backbones such as ESM-2 and discriminant features for the biological properties or fitness of the parsed protein sequence.

The paper introduces EvoCircuit, an active-learning framework that exploits patterns obtained from cross-layer transcoders (CLTs) to integrate feature-relevant or fitness-related signals into the candidate selection procedure, integrating the model’s internal representations into the active learning loop.

The CLTs are instantiated along the layers whose activations we want to inspect, and the sequence-level representation is computed as the average activation at position p of sequence x.

Once the average representation is computed, a candidate pool of CLT features is constructed by retaining features whose activations vary across the measured sequences and sampling them uniformly within each layer using balanced quotas. Each candidate feature is then evaluated by its Spearman correlation with measured activity and the stability of its correlation sign under bootstrap resampling. Features meeting both thresholds are retained, keeping at most K features with the largest absolute correlations. These selected features define a task-adaptive distance used to retrieve candidate variants near the highest-activity variants measured so far.

In the end, EvoCircuit uses the coordinates to create a distance measurement between the candidate sequences and the experimentally validated high-activity variants, finally combining the resulting ranking with the predictor's inference.

EvoCircuit uses feature-guided retrieval to complement rather than replace the activity predictor. While one portion of the batch preserves globally high-scoring candidates, the other is selected from regions identified through activity-associated internal pLM features.

Regarding sequence mutation gains, the experiments focus on single-mutant deep mutational scans, where the model demonstrates marginal improvements on a subset of baseline use cases, and combinatorial mutations, where the model still shows limited improvement over the baselines. The authors also provide further ablation experiments to better motivate the method.

Critical strengths

The paper is overall easy to read and clear. The problem and the background is well introduced and motivated, highlighting the links between the AI applications and the causality observations coming from mechainstic interpretability applications. The experimental section is well organized and the authors provide a thorough ablation and discussion of the limitations and the observations of the method itself.

Critical weaknesses

The method is overall well-partitioned, but some parts are overcomplicated with discontinuous and poorly formalized notation. It is also not clear from the main content what the algorithm itself is and how CLTs intervene in the optimization pipeline. The figure in the Appendix should be included in the main paper to improve clarity and should be better explained. The results shown in the experimental section are not sufficiently exhaustive. The authors should include more benchmarks and further mutational analysis to better validate the generalization capabilities of the method. Moreover, the gains obtained by applying the method over the single-mutation case are not even substantial across all proposed configurations. This demonstrates a lack of evidence for the effectiveness of the method and a lack of analysis regarding the reasons behind the reported results.

Questions for authors

a. The proposed experiments only look at single mutants; perhaps adding higher-order mutants could have shown better gains from the insertion of the EvoCircuit feedback. Why didn't you include those? Is there any observation that limits the application of such a method?

b. Figures 1a and 1b are not clear in their current form. Rearranging them or providing a tabular version would help, along with adding more rigorous commentary focused on the quantitative comparison against random CLT.

c. The ablation setting is not clear and is somewhat confusing. Better formalization is needed, along with auxiliary figures or tables to clarify the results.

d. In Section 3.1, the sequence-level descriptor is purely done by averaging per position, without weighting the average by any signal. Is there evidence that this is better than applying a per-position weight?

e. Section 3.2 screens the pool based on a smaller dataset. Is there any evidence of bias in the screening coming from the dataset? Does the choice of data affect generalization in any way?

1 comment
AuthorsOpenReview ↗

Method Clarification, Expanded Evaluation, and Feature Selection Analysis

Thank you for your thorough review and detailed comments.

[Q-(c) Method clarity and pipeline]

We revised Sections 3.1–3.3 and Figure 1 to clarify the acquisition procedure. Section 3.3 explains how feature-guided retrieval and predictor-based ranking are combined to select variants for the next round.

[Q-(a) Evaluation beyond fixed-site benchmarks]

We added Pab1, GFP, cgreGFP, and AAV to evaluate libraries with variable mutation-site subsets (Sections 4.1 and 5.2, Table 4). Pab1 contains single and double mutants; the other three also include higher-order variants. The evaluations cover the full measured pools, rather than only higher-order variants.

Each comparison uses the same predictor configuration, 10 paired seeds, 64 shared uniform-random initial measurements, and 112 additional variant selections. EvoCircuit increases mean top-10% discoveries in 17 of 20 comparisons against greedy acquisition. On cgreGFP/FolDE, discoveries increase from 20.5 to 29.9 (+45.9%). These results show that the discovery benefit extends beyond the fixed-site benchmarks.

[Q-(b) Figure interpretation and feature-selection effects]

Figure 1(b) in the submitted version showed mean gains over matched random CLT features, not total discoveries. Revised Figure 2 shows absolute counts and separates full-campaign results from fixed-state comparisons.

For RF under unscreened initialization, mean top-10% hits among eight feature-guided selections increase from 1.7 to 2.2 (paired gain 0.5; pointwise 95% CI [0.3, 0.7]), averaged over saved rounds and seeds across four datasets. Because the observed history, predictor, anchors, and budget are fixed, this comparison supports the contribution of activity-based feature selection to discovery among feature-guided selections, rather than estimating cumulative campaign gain. Section 5.3 explains this distinction, with paired uncertainty reported in Appendix H, Table H.2. Complementing this acquisition evidence, Section 5.4 shows how the CLT features used in retrieval can be interpreted through their activation profiles and removed to test their influence on individual recommendations.

[Q-(d) Uniform versus mutation-local pooling]

Uniform pooling summarizes each feature without retaining its activation locations. We experimentally evaluated a label-free alternative that assigns weight two within ±5 residues of a mutation and weight one elsewhere, then normalizes. Both arms use EvoCircuit with RF, 10 paired seeds per dataset, 64 shared initial measurements, and 112 additional variant selections. This comparison and the diagnostic in (e) are reported here, not in the current revision.

Mean top-10% discovery AUCs and paired gains with pointwise 95% CIs are:

DatasetUniformWeightedPaired gain [95% CI]
GCK11.413.2+1.8 [1.3, 2.4]
GFP16.817.6+0.9 [-0.6, 2.2]

Paired gains are computed before rounding. Although the GFP difference remains inconclusive, the GCK result shows that mutation-local weighting can improve acquisition within the same framework. Uniform pooling is a reference configuration, not a requirement of EvoCircuit or a universally optimal choice.

[Q-(e) Small-sample stability and acquisition utility]

Bootstrap sign stability (Section 3.2) is a screening criterion for consistency under resampling of observed measurements, not a guarantee of generalization. We therefore assess the selected features beyond their bootstrap scores.

In an additional diagnostic, we evaluated selected-feature correlations on 256 candidates not yet selected for evaluation per saved state (GFP and Pab1; five seeds per dataset; rounds 1 and 7), using their activity labels only for evaluation. Mean absolute correlations decrease from 0.4–0.5 on measured variants to about 0.2 on these candidates.

The acquisition evidence comes from the fixed-state controls in Section 5.3, discussed in (b): activity-selected features yield more new hits than matched random features under the same history and predictor. Although these controls do not isolate the bootstrap filter, they support the usefulness of the selected feature sets for acquisition in the evaluated settings.

Together, these results clarify the acquisition procedure, extend its empirical evaluation, and support activity-based feature selection as a useful complement to predictor-based ranking. We welcome further questions about the method or analyses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.