acceptodds
Under review as a conference paper at ICLR 2027

GFlowNets build feasible designs for directed evolution campaigns

Abstract

In a directed evolution campaign the budget, not the size of the search space, is the binding constraint: a few hundred assays, a plate of variants a round, to improve an enzyme or a binder. When a hard feasibility constraint applies, the pipelines used today — a genetic algorithm, machine-learning-guided design (MLDE), CMA-ES over a continuous relaxation — cannot represent which sequences are buildable; they draw candidates and subsequently reject the illegal ones, which wastes budget on unmakeable variants and, when feasible designs are scarce, leaves the method unable to run at all. Building each design move by move under a mask that permits only legal continuations removes the bottleneck entirely, since every proposal is then feasible by construction. We present EvoGFN, a GFlowNet that constructs designs over the construction graph rooted at a campaign's parent, enforcing feasibility by masking each move, and test it at wet-lab budgets against those pipelines. Its mean fitness is the highest on every task in our synthetic suite, decisively where few sequences are feasible, though a CMA-ES baseline reaches the same ceiling on the task where both saturate; on the two empirical landscapes it leads narrowly. Throughout it proposes only feasible designs, where the pipelines it is compared against spend much of their budget on designs that cannot be built. The same advantage holds on a real, structure-derived constraint, a folding-stability predicate on the GB1 landscape thresholded on predicted ΔΔG, where a proximal baseline matches it on fitness only at the tightest threshold. Controls separate mask from policy and trace the dependence on predictor quality: porting the mask to a plain genetic algorithm gives it full feasibility and a fitness gain, untraining our own policy costs it little except where the feasible set is small, and injecting noise into the fitness predictor that ranks candidates before any assay is spent erodes the lead only where that predictor was informative. The advantage narrows where the feasibility predicate is simple enough to admit an exact projection — a dynamic program that recovers the best design under an additive decoding score, though not under the true fitness — against which the learned sampler keeps a small edge on some tasks and ties on others. A masked sampler gains most where the constraint is strict, and where an exact decoder can be built it closes much of that gain. Where the constraint couples every position, a local mask reaches almost none of the legal designs, and we leave that regime as an open frontier.

Primary area:
generative models
Keywords:
directed evolution, protein sequence design, gflownets, constrained generation, masked construction, active learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Ef6qOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies protein design when only a limited number of experiments can be run and candidate sequences must satisfy known requirements. The authors propose EvoGFN, a GFlowNet that starts from an existing sequence and changes it one position at a time. Each change is checked against the requirements, so candidates that fail them are not generated. A fitness predictor trained on available observations guides candidate generation. The evaluation uses synthetic tasks, existing GB1 and TrpB measurements, and GB1 with an added predicted stability requirement. The authors also test untrained generators, larger candidate pools, and other methods with the same feasibility checks to understand which parts of the approach produce the gains.

Critical strengths
  1. Gains are obvious and cannot be explained simply by baselines producing more infeasible candidates. Under the mild stability requirement in Table 2, EvoGFN achieves a normalized best fitness of 0.795, compared with 0.690 for PEX and 0.685 for PEX-ensemble, despite their feasibility rates already being 99.6% and 99.9%. EvoGFN also outperforms masked-ALDE and masked-MLDE, which both achieve 100% feasibility. This is meaningful evidence for the effectiveness of the complete workflow.
  2. The analysis of the method’s limitations is useful and includes targeted controls. The paper goes beyond reporting best fitness: it explains how local checks can exclude valid final sequences and studies completion-aware alternatives. The untrained-generator, larger-pool, and exact-decoding controls also show an effort to understand where the gains come from rather than only presenting favorable rankings. These analyses help establish the method’s scope.
Critical weaknesses
  1. The additional benefit of GFlowNet training is not sufficiently isolated, leaving the main methodological contribution unclear. The mild-gate results in Table 2 establish an advantage for the complete workflow, but do not directly show whether the gain comes from learning the generator, the candidate distribution, the number of candidates, or the prediction and selection steps. The masked-ALDE and masked-MLDE baselines control feasibility, but they do not differ from EvoGFN only in whether its generator is trained.
  2. Table 3 provides a more direct trained-versus-untrained comparison, but its GB1 and TrpB tasks have no stability constraint and therefore do not directly explain Table 2. It seems that predictor corruption affects both generator training and candidate screening, and the experiment cannot separate the two effects.
  3. The budget accounting in some main comparisons may overstate the practical assay-efficiency advantage. EvoGFN’s feasibility checks are treated as computation, whereas infeasible candidates from several baselines consume assay slots without returning fitness labels. Since the rule is known in advance, it is unclear why these methods cannot discard invalid candidates computationally before spending an assay. This has substantial consequences: ordinary MLDE never obtains enough usable data to start training on the three main constrained tasks.
Questions for authors
  1. Can every method apply the same feasibility check before an assay? Do the main advantages in Figure 1 remain when every method can prefilter invalid candidates, with computational costs recorded separately?
  2. How do EvoGFN, PEX-ensemble, and the two masked baselines in Table 2 differ in candidate generation, prediction models, candidate counts, and final selection, and how these differences support the central methodological claim?
Reviewer_q6p5OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose a GFlowNet method for directed evolution, which allows to encode feasibility constraints directed within the generation process. Under specific feasibility metrics, they show on both a toy example and on a computational experiment with wetlab-size budgets, that the proposed GFlowNet-based method allows to achieve better fitness-diversity results than several strong baseline methods.

Critical strengths
  1. The paper explores the intriguing idea of using GFlowNets in ML-assisted directed evolution campaigns. This is original to my knowledge, and might lead to interesting new viewpoints and methods.
  2. The baseline methods considered are relevant and well-chosen, although it is possible their inputs (e.g., the reward function) has not been properly set to assess the proposed methods in a convincing manner (see weaknesses).
Critical weaknesses
  1. (Main concern 1 - Conceptual) To the best of my knowledge in ALDE and in general protein design, people already regard by fitness the combination of functional activity and any hard-constraint one might want to impose (i.e., if a protein does not satisfy them, then its fitness is usually considered to me 0). This does not seem well-aligned with what is written in sec. 1 and later. Moreover, to my knowledge in most cases for proteins one (i) does not require a synthesis route alike small molecules, and (ii) computational classifiers (e.g., for foldability) are often unreliable, and therefore it is not true that “for building real proteins a folding classifier that must call the variant expressible” as written in Sec. 1. As it is possible that such learned classifiers provide false negatives (and false positives), especially OOD, where is most relevant for many discovery targets.
  2. (Main concern 2 - Experiments) Since the authors define a specific, arbitrary notion of validity, it would be essential to test as baselines existing methods where the fitness (the reward) is simply re-defined by multiplication with an indicator function capturing validity according to the new concept, to assess whether a trivial change in reward function can solve the problem tackled in this work. My current belief is that it typically does.
  3. (Main concern 3 - Conceptual Framing/Goal) The (statistical) budget constraint mentioned in the first line of the abstract should regard the wetlab budget, not computational. To my knowledge, sub-selecting more promising samples according to any knowledge/information about their “validity” or investing more (computational) budget into fine-tuning methods based on such accessible knowledge is not directly related to the mentioned wetlab bottleneck (i.e., the statistical budget/cost). This connection appears to be quite unclear at the moment in this work as the term “budget” seems used to express both costs. Moreover, since typical expression wetlab experiments take weeks (e.g., 2 weeks), in my understanding, computational budget is really not a bottleneck in practice.

Regarding the low presentation score (not critical for assessment):

  • Abstract is very badly written, hard to parse, with multiple super long and complex sentences (e.g., at line 28 it starts a sentence that is 6 lines long, and very hard to understand), typos (e.g., “policy costs it little”?), and employing terms and concepts never introduced (e.g., “a dynamic program that recovers the best design under an additive decoding score”?). Writing is overall better in the majority of next sections.
Questions for authors

see weaknesses

Reviewer_wms8OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

EvoGFN is a GFlowNet-based algorithm designed for directed evolution campaigns that constructs candidate protein sequences step-by-step while applying a local mask to ensure every move leads to a feasible design. By filtering out unbuildable variants before proposing them, it avoids wasting limited wet-lab assay budgets on rejected candidates. Compared to traditional pipelines, EvoGFN consistently achieves superior or competitive fitness and higher candidate diversity across synthetic and biological benchmark tasks.

Critical strengths
  • Good problem setting to include feasibility as the criteria while sequence design.
  • Shows a good improvement on the feasibility of designs.
Critical weaknesses
  • Missing Baselines: Key biological sequence optimization baselines such as Feedback GAN (Gupta & Zou, 2019) and IDEAS (Pandey et al., Interpretability Driven Evolutionary Approach for the Design of Biological Sequences, 2026) are missing. Authors claims are very strong: ". Our sampler leads every published pipeline on every task in our synthetic suite"

  • Why do authors runs only 4 design iteration. I believe should be increased to 10.

  • In figure 1, for ehrlich-feas, how is the fitness scores shaded region going above 1? Can the fitness value be more than 1?

Questions for authors

Same as weaknesses.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the revision for the Area Chair and reviewers

Dear Area Chair and reviewers,

Thank you for the comments, which helped clarify the framing and strengthen the evaluation. The revision tests whether incorporating a supplied feasibility rule into a learned generator improves the fitness reached under a fixed assay budget when competing methods can apply the same check before an assay. Computational screening and assay use are counted separately. The main findings are as follows.

  • Fitness gains with pre-assay filtering. Against the best filtered baseline, EvoGFN reaches normalised best fitness of 0.574 against 0.365 on the ALDE protocol, 0.447 against 0.369 on EvolvePro, and 0.795 against 0.704 at the mild GB1 stability gate (Tables 1–2). In these comparisons, rejected proposals consume no assay slots: the differences concern which feasible designs are selected for the same assay budget. At the strict gate, filtered PEX variants lead (0.920–0.921 against 0.901), and several adapted methods reach the same ceiling on sparse Ehrlich. The revised claims identify these settings explicitly.
  • Evidence for generator training. On both protocols and both stability gates, training improves over the same untrained generator by 0.04–0.10 at matched proposal counts and by 0.02–0.05 when the untrained generator receives four times the proposals, in the paper's normalised fitness units (Table 3, Section 3.4). These comparisons match the mask, predictor architecture and selection procedure. A separate frozen-history control holds the fitted predictor and measured history fixed, and favours training at both gates and both pool sizes (Appendix F.3). The genetic-algorithm control shows only a small additional gain from masking over pre-assay filtering, so we do not attribute the full workflow advantage to masking alone.
  • Broader comparisons and longer campaigns. IDEAS and an adapted FBGAN join the six original baseline families, with pre-assay filtering variants. Selected baselines also receive zero-label, feasibility-ranking or masked-generation controls; their configurations are specified in Appendix F.4. Extending the two protocols, GB1 and TrpB to twelve rounds shows that EvoGFN's lead persists on the protocols, while adapted methods overtake it after round 9 on GB1 and TrpB (Figure 2). On the tested Ehrlich instances, an exact decoder optimises an additive positional score over the feasible Hamming ball at about a hundredth of EvoGFN's CPU cost; EvoGFN retains an average fitness margin of about 0.03 across the tested cells (Section 3.6).

Meaning of the biological constraint. In GB1, fitness is measured binding, while feasibility means passing a separately supplied predicted-stability threshold. A variant can bind well and fail that threshold. This setting isolates how a generator uses an available acceptance rule. Without measured stability labels across the library, we use a simulated, ThermoMPNN-based gate and explicitly distinguish compliance with it from evidence that a protein folds or expresses. Corrupting the mild gate with false rejections, false acceptances, or both leaves EvoGFN's normalised best between 0.79 and 0.82, against 0.795 with the reference gate (Section 3.3); this tests sensitivity to those perturbations, without validating the gate's biological accuracy.

Discussion and scope. The prefiltered comparisons address the concern that gains arise solely from charging assays for computationally identifiable failures. They do not settle how much further the filtered baselines could improve with more search, or with q6p5's proposed combination of zero-label training and free rejection, which remains untested. Ef6q noted on 15 November that the additional experiments largely address the original concerns and asked what drives the difference from PEX-ensemble. Appendix F.4 now describes that workflow, and our closing reply distinguishes the shared predictor architecture from the generation and selection differences that remain. We did not receive a reply from wms8 to our 9 November comment and have summarised the completed additions in a closing reply.

The abstract and introduction have been rewritten, the broad "leads every published pipeline" statement removed, and the campaign steps and appendix explanations clarified. We hope the revised comparisons and controls help the committee assess the contribution at the scope supported by the results.

Thank you for your time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.