acceptodds
Under review as a conference paper at ICLR 2027

EGRET: Exploring Generalist and Specialist Repertoires via Posterior Lineage Sampling over Epistatic Trees

Abstract

Machine-learning-guided protein engineering usually optimizes one sequence at a time, although related conditions may be better served by descendants that share early mutations and specialize after branching. We introduce _assay-limited generalist-to-specialist diversification_: given a common scaffold, a finite sequence library, and limited sequence–condition measurements, a method jointly selects a condition niche and a rooted mutation tree under a shared construction budget. The objective measures how much accessible specialist headroom the tree captures beyond the best reachable generalist. This decision couples specialists by shared lineage cost and exposes condition-wise surrogate maximization to the optimizer's curse. EGRET uses a hierarchical epistatic posterior and posterior lineage sampling for assay allocation. It commits a final tree using measured leaf–condition utilities, providing a verifiable lower bound on its value. Diverse experiments evaluate repertoire quality, explanation faithfulness, and decision traceability across protein-engineering scenarios with varied objectives and assay budgets.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
biological sequence design, decision-focused learning, active learning

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

EGRET is a protein engineering framework that designs affordable families of specialized variants by organizing them into a rooted mutation tree with shared early edits. It uses a structured epistatic model and posterior lineage sampling to prioritize lab tests for sequences that consistently support high-value trees across uncertain landscape draws. Rather than relying on predictions, EGRET commits its final specialist repertoire using verified lab measurements to guarantee a verifiable lower bound on performance.

Critical strengths
  • Novel Problem Formulation: The paper formalizes a realistic biological task (assay-limited generalist-to-specialist diversification) that bridges multi-objective search and cost-aware sequence design by evaluating mutation trees with shared construction steps.

  • Methodological Rigor and Guardrails: EGRET effectively counters the optimizer's curse by coupling structured posterior lineage sampling for assay allocation with a measured-value commitment step that provides a verifiable lower bound on repertoire value.

  • Comprehensive Empirical Evaluation: The method is rigorously benchmarked against nine strong baselines across 20 diverse biological datasets, supported by extensive ablations, intervention studies, and explanation faithfulness audits.

Critical weaknesses
  • EGRET only works on the datasets with pre-alinged set of protein sequences. can author throw light on how will EGRET work on dataset with just random set of sequences (AAV fitness, Binding affinity datasets)?
Questions for authors

Same as weaknesses.

Reviewer_SJFhOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies assay-limited generalist-to-specialist diversification: selecting a condition niche and a family of sequence variants that can be constructed through a shared mutation tree. The objective measures the repertoire’s coverage relative to the best reachable generalist under the same construction budget. EGRET combines a hierarchical site/residue/pair model, posterior sampling followed by feasible tree search, and adaptive allocation of sequence–condition measurements. Final selection uses measured coverage, giving a utility lower bound under the noiseless assay model. The main contributions are the joint repertoire-design formulation and the associated acquisition procedure. Validation covers twenty finite-library benchmark instances, comparisons with optimization and active-learning methods, and extensive ablation, robustness, calibration, and explanation analyses.

Critical strengths
  1. The problem formulation is useful and well motivated. Individually strong variants may be difficult to construct together under a shared budget, and the proposed objective makes this dependence explicit. Separating assay costs from construction costs and comparing against a budget-matched reachable generalist gives the problem a clear interpretation. I find this a meaningful contribution beyond optimizing individual endpoints or returning a diverse candidate set.

  2. The complete system has strong empirical results. Table 1 reports mean headroom capture of 0.641 and the highest mean performance on 17 of 20 instances. The additional comparison under a common measured-value commitment rule is particularly helpful: EGRET reaches 0.670 versus 0.519 for DADO-refit on the sixteen-instance subset (Table 17). These results support the usefulness of construction-aware acquisition. The paper also reports failure cases, including loss of the advantage under strong observation noise and destroyed mutational locality, which helps establish the method’s scope.

Critical weaknesses
  1. I had difficulty reconstructing the probabilistic model actually used by the algorithm. Section 4.1 derives a joint hierarchical Gaussian posterior, while Appendix B.2 uses two sequential fits and an approximate covariance. The appendix acknowledges this approximation, but the joint sampling procedure across conditions and shared unseen substitutions remains insufficiently specified. This matters because optimal-tree probabilities depend on joint correlations, not just marginal variances. The relationship between the working sampler, the calibration diagnostics, and Proposition C.2 should therefore be clearer. At present, I can follow the reference model more readily than the implemented acquisition distribution.
Questions for authors
  1. Do the authors have a matched comparison between EGRET and EGRET-mean across the full benchmark, keeping the surrogate, candidate pool, tree solver, assay allocation, and commitment rule fixed? How does the proposed acquisition compare with single-draw tree Thompson sampling or PS-BAX under those same settings? These results would materially affect my assessment of the methodological contribution.

  2. What joint covariance defines the implemented landscape draws? In particular, how are conditions coupled, and is an unseen substitution assigned the same random effect across sequences containing it? Are sampled utilities constrained to ([0,1]), and which parts of Proposition C.2 apply to the working approximation?

Reviewer_rfucOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2(+1)
Soundness
1
Read review
Paper summary

The paper considers learning guided protein engineering, in particular finding high performing sequences xx from a finite library XX under $c \in C, M_{full}=| C | $ related conditions,formalized as conditional utility f:X×C→[0,1]f:X \times C \to [0,1]. Starting from an initial sequence x0x_0 the paper considers a XRX_R as the reachable set via mutations and other prefix sharing incremental modifications of x0x_0 with  construction budget RR and an assay budget BB for evaluations of sequences.

The  paper proposes to consider the set of constructible search trees reachable from the initital sequence x0x_0 within the budget RR ,  t∈TR(x0)t\in T_R(x_0) and the concept of a niche C’⊆C,M=∣C′∣≤MfullC’ \subseteq C, M=|C'| \leq M_{full}, and to consider the normalized regret ΦR(C,T)≔V(C,T)−GR(C)UR(C)−GR(C)\Phi_R(C,T)\coloneqq \frac{V(C,T)-G_R(C)}{U_R(C)-G_R(C)}  where V(C′,t)=1/M∑c∈C’max⁡x∈δtf(x,c)V(C',t)=1/M \sum_{c \in C’} \max_{x\in \delta t f(x,c)}  is the average condition specific top performance amongst the leaf candidates of tt (δt\delta t being the leaves of tt ) , Gr(C’):−max⁡x∈XR1/M∑c∈C’f(x,c)G_r(C’)\coloneq \max_{x\in X_R} 1 /M \sum_{c\in C’} f(x,c) being the best reachable generalist across the niche, and UR(C’):−1/M∑c∈C’max⁡x∈XRf(x,c)U_R(C’)\coloneq 1 /M \sum_{c\in C’} \max_{x\in X_R} f(x,c) being the unrealizable upper bonud of performance within XRX_R allowing to switch to the specific champion per condition cc. The property is motivated for being of interest due to e.g. enzyme  construction via directed evolution.

The authors then create a hierarchical linear model of the utility f^(x,c)=z(x)T(θˉ+δc)⏟θc+ϵxc;ϵxc∼N(0,σc2)\hat{f}(x,c)=z(x)^T\underbrace{(\bar{\theta}+\delta_c)}_{\theta_c}+\epsilon_{xc};\epsilon_{xc}\sim \mathcal{N}(0,\sigma_c^2)  with θˉ\bar{\theta} being shared across conditions and δc\delta_c being a condition specific term.

(I will be more vague now, see weaknesses section for why)

The authors state a posterior that they aim to calibrate such that it allows transfering “a sites mean to an unseen residue, but not the site’s confidence”.

The method then estimates the probability πC(x)\pi_C(x) that a sequence xx participates in the best tree for a draw of the posterior model of f^\hat{f}   and the score  aC(x)a_C(x) which tracks smoothed counts of a sequence being in the leaves of the top JJ leaves or being in the top scoring generalist set.

There is consideration for reachability and cost aware tree construction discussed, as well as what the paper calls “spectral assay aperture and adaptive allocation” and verified commitment”.

The method is compared with 9 baselines including submodular tree orienteering (STO),BOCS and other bayeisan combinatorial optimization methods, as well as a random control. Various ablations and further studies are performed.

Critical strengths

Post rebuttal update: the papers key idea appears promising, and the benchmarks, if they hold up and are meaningful, are very impressive.


None

Critical weaknesses

Post rebuttal update: the authors have substantially improved the clarity of présentation throughout rebuttals and provided further answers and clarification in the comments, however I retain substantial uncertainty and doubts as to the soundness is the evaluation and real world applicability, mainly due to the still quite challenging to read presentation and lack of information and context (see end of the discussions with authors)

The paper is not intelligible by reading the main body, due to excessive vibe coding.

  1. The main body makes it impossible to understand how EGRET actually works. no pseudo code or algorithmic listing is given in the main body (it is in the appendix, but reviewers are not required to read the appendix, and  there are 30 pages of appendix), nor a high level summary specific enough to analyize the method,  indeed the use of πC\pi_C or aCa_C are never explained on how they are used to select sequences or how they get the  trees they are derived from.  And 235 onwards is even worse of non-coherent info dumping than the sections until then.

    1. more specific points of illegibility/LLM speech:

      1. 037ff:  Sentences start with inanimate nouns and concepts verbing, and are rapid fired without explanation for a conference where the audience might not understand wetlab terminology (e.g. promiscous activity will have a different connotation outside of a lab)

      2. 036: why do these advances motivate the design question? nonsequitor

      3. 050 to 060: nonsequitor dump of work loosely related, with no clear point for the paper. same papers as are alter repeated in the related work section

      4. 061 to 067: tech jargon dump with no informatoin content due to lack of precision

      5. all claims are simply stated, with no throughline,and all meaty content is deferred to appendix, without a proof sketch

      6. there is so much more but listing everything would be fixing the paper which is not my job

  2. Ignoring this, the evaluation is gimped, of course EGRET will perform well according to metrics defined in the paper, at minimum the baselines performance metrics would need to be defined as well

  3. The motivation is insufficient for why we would care about the metric defined: construction is rarely the bottleneck https://www.frontiersin.org/journals/bioengineering-and-biotechnology/articles/10.3389/fbioe.2015.00135/pdf but rather the testing of the constructed sequences. ALDE and the datasets the paper builds on for evaluation are possible because of how cheap it is to construct sequences (with the cost being the assay).

  4. even if all these problems didn’t exist, the notion is ill formed: the algorithm can pick both conditions and sequences, meaning that it it can simply ignore harder conditions and focus on easier ones, which defeats the point

  5. from what I can tell, the algorithmic meat is a remix of the citation (STO mainly) which would be fine if cleanly presented as an integration and legible report, but not in context

Questions for authors

None. Please write your paper by hand, properly motivated and written, the notion could be interesting.

Post rebuttal update: I thank the authors for substantial engagement

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Updates and Acknowledgments

We are sorry that it is a weekday. So this revision takes time, but we are truly putting our hardest human effort into connecting all the requirements. It is almost pushing us to our limits and taking up all our nighttime, but in that way, we learn. We hope this last-minute version can nicely address the presentation points raised by Reviewer rfuc (and we hope you will reconsider your rating in the next phase, if possible). The blue text provides evidence of our improvements.

We have added a concise motivation, mathematical background, and biological context to the Introduction and reorganized Appendix A and Related Work to make it easier for everyone to understand the intuition and real applications behind our problem. We have also improved the presentation of the Experiments, that is, moving some tables and figures to Appendix E and keeping only what is really important in the main text, as suggested by Reviewer rfuc. Regarding the Methodology, Section 4 could be read independently without the Appendix, Appendix B just provides low-level implementation details, and we hope this can convince Reviewer rfuc.

These changes answer all the questions: Why output a tree? What is construction for? Why specialists? Which real-world use cases need this? What are the intuitions behind our design? They are now explicitly explained.

We hope our paper is now ready for this edition of ICLR. There are some aesthetic issues in the Appendix that we did not have enough time to fix, and we will address them in the camera-ready version if this paper still has a chance.

We hope that the strengths of our paper will be recognized in the decisions of the AC and reviewers, and by the readers: (1) a well-motivated problem formulation with potential real-world applications, (2) a novel and explainable method, and (3) extensive experiments across multiple scenarios, with all results reproducible through our publicly available source code.


I would like to say one last thank you to Reviewer rfuc and acknowledge all the reviewers in this version to express our gratitude for your detailed comments and kind assistance, especially the time and effort of Reviewer rfuc.

We also hope that the AC can recognize Reviewer rfuc for their contributions to the scientific community and their expertise in discussions, which help advance the field. We have learned a lot from them.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.