acceptodds
Under review as a conference paper at ICLR 2027

Mutation Flows: Decomposable Variable-length Flow Matching with Block Insertions

Abstract

Biological sequences evolve through local edits: single-residue substitutions, deletions, and block insertions of several residues at once. Discrete flow matching is a natural fit for this generative process, but existing formulations fall short in three ways. They fix sequence length or restrict edits to single tokens; they cannot decompose the joint prediction of edit type, block length, and token content; and they fix the source-to-target sequence distance implicitly through the choice of coupling, leaving it uncontrollable at inference time. We introduce Mutation Flows, which addresses all three. (i) Block insertions of arbitrary length are modelled by a zero-truncated Poisson distribution and resolved in a single transition. (ii) An exact planner-denoiser decomposition splits the objective into two losses with disjoint parameters, whose denoiser term reduces to a standard fixed-length discrete flow matching objective, admitting any pretrained biological sequence model as a drop-in denoiser. (iii) Divergence conditioning allows for coarse-grained control of the sequence divergence from the source sequence during inference. This approach enables flows between the same source and target distributions without degrading into a trivial mapping, as is the case for optimal transport. On antimicrobial peptides, Mutation Flows stays closest to the data distribution throughout the denoising trajectory, remains effective with a pretrained ESM-2 dropped in as the denoiser, preserves these gains in the few-step denoising setting, and achieves state-of-the-art performance in designing active peptides.

Primary area:
generative models
Keywords:
discrete flow matching, peptides, biological sequences, variable-length

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 4 reviews, above 27% of the others.
Reviewer_hrJjOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors introduce Mutation Flows, a variable-length discrete flow-matching method supporting substitutions, deletions, and block insertions. They sample insertion lengths from a zero-truncated Poisson distribution and construct training paths over aligned sequence pairs, resolving each insertion block as one event. A planner predicts edit locations, operation types, and insertion lengths, while a separate denoiser predicts residue identities. The authors derive a planner-denoiser loss decomposition, claim equivalence of the denoiser objective to fixed-length discrete flow matching, and condition generation on a requested divergence between independently sampled source and target sequences. They trained on 871,918 AMP sequences of 10-50 residues and evaluated intermediate-state quality, generation with few sampling steps, and predictor-guided optimization. Comparisons included Edit Flows, FlexMDM, several AMP generators, and a frozen ESM-2 denoiser. The full model achieved lower embedding-space MMD and higher predicted AMP hit rates, although competing methods achieved better scores under the independent APEX activity predictor.

Critical strengths
  • The ability to insert a complete block in one transition could be useful when a design requires adding a motif or changing a loop length, and it provides a concrete extension to single-token edit flows.
  • The frozen ESM-2 experiment offers practical evidence that the planner can work with an existing sequence model. Reusing pretrained denoisers could reduce the cost of adapting the approach to other biological sequence families.
  • Evaluating intermediate sequences is useful for guided search, where predictors score candidates before generation finishes. The authors also examine small sampling budgets, addressing a practical limitation of iterative sequence design.
Critical weaknesses
  • My main concern is that the experiments do not isolate the methodological advance over Edit Flows, FlexMDM, and planned denoising. Appendix B states that the conditioned model uses Needleman-Wunsch alignments while the unconditioned control only pads sequences. The planner-denoiser comparison also changes the insertion-content model from a shared distribution to position-specific predictions. Meanwhile, the single-model block-insertion variant performs worse than Edit Flows in Table 1. I would need cleaner ablations to identify what the block transitions themselves contribute.
  • The claims of exact decomposition and denoiser equivalence need much stronger justification. Different edit events can produce the same destination sequence, which affects the sequence-level rate and its logarithm. The denoiser also receives different contexts during training and sampling. I can't yet get these details with the stated proofs, becase the exactness claims are central to the contribution.
  • The AMP experiments do not establish the need for this general variable-length framework. Short peptides and predictor-based activity scores provide limited tests of structured block editing, and low MMD is difficult to interpret when starting sequences already come from the target distribution. The activity advantage is also mixed: guided Mutation Flows has an APEX-predicted MIC of 80.8 uM, compared with 56.2 uM for Hyformer. Protein and DNA tasks with meaningful insertion constraints would make the contribution much more convincing.
Questions for authors
  1. Could the authors isolate block insertion by comparing the full planner-denoiser model with an otherwise identical single-token-insertion variant? Please hold alignment, divergence conditioning, position-specific content prediction, and sampling budget fixed, and give Edit Flows and FlexMDM the same conditioning information.

  2. Inserting A after either residue of AA produces AAA, so the sequence-level transition rate must sum both event rates. How does the logarithm of that sum separate exactly into planner and denoiser terms? Please clarify whether the generator acts on sequences or labelled edit events, and explain how Eq. 21 handles substitutions that retain the original residue.

  3. Algorithm 1 builds the denoiser input using all remaining target edits and removes target deletion positions, whereas Algorithm 2 applies only the current sampled plan. Please justify Proposition 2 under these different contexts and the correlated masks produced by block insertions. Does the equivalence concern the complete training distribution or only the cross-entropy form?

  4. Does the network receive time tt separately? Appendix B gives only bt=(1−κt)db_t=(1-\kappa_t)d, the requested divergence dd multiplied by the remaining schedule fraction 1−κt1-\kappa_t, whereas Proposition 3 conditions on both time and divergence. Different (t,d)(t,d) pairs share the same btb_t. Please make sure the implementation and the proof agree.

  5. Since source sequences already follow the target distribution, please include unchanged-source and corpus-resampling controls. Report actual source-to-output divergence η(x0,xout)\eta(x_0,x_{\mathrm{out}}) against the requested value, and measure novelty against the entire training corpus. Absence from a 2,000-sequence reference set does not establish novelty relative to training.

  6. Algorithm 2 uses no-edit probability 1−hΛi1-h\Lambda_i, where hh is the step size and Λi\Lambda_i is the total edit rate at position ii. How is this kept nonnegative with two or four sampling steps? Please report any clipping and measure wall-clock latency and actual model calls, since the reported NFE counts sampling steps.

  7. Could the authors add a longer-protein task involving loop or motif insertions and a DNA regulatory-sequence task involving motif placement and spacing? Please assess functional retention or improvement at matched edit distances and length changes. These settings would directly test whether block editing provides a useful capability.

1 comment

I look forward to the response from the authors and am willing to revise my score with clarifications to my questions.

Reviewer_X1gfOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

Existing discrete flow matching methods are typically limited to fixed-length sequences. Building upon EditFlow, this paper introduces block insertion, planner-denoiser decomposition, and divergence control to enable more flexible and controllable variable-length generation. In AMP generation experiments, the proposed method achieves lower ESM-C embedding MMD than baselines such as EditFlow and FlexMDM, along with competitive predicted antimicrobial activity.

Critical strengths
  • The method demonstrates some novelty. It introduces block insertion into variable-length discrete flow matching. The planner–denoiser decomposition is well motivated from a probabilistic modeling perspective and provides an exact objective factorization. It also supports integration with pretrained masked PLMs such as ESM.
  • The experimental results are competitive. The method significantly outperforms major baselines in trajectory MMD and few-step generation. Experiments with frozen ESM-2 demonstrate the feasibility of using a drop-in denoiser. It also achieves strong predicted activity, diversity, and novelty in AMP generation and optimization.
Critical weaknesses
  • Insufficient validation of the contribution of block insertion. Although multi-residue block insertion is highlighted as a key contribution, the paper lacks analyses of insertion length distributions, biologically meaningful motifs, and the independent performance gains over single-token insertion.
  • Confounding factors in the evaluation of divergence conditioning. The conditioned [C] and unconditioned [U] variants use different alignment strategies.
  • Limited biological validation. Experiments are restricted to AMP generation, without evaluating generalization to general protein sequence generation, protein engineering, or other biological sequence tasks.
Questions for authors
  • Could the authors provide further analyses of insertion block lengths and generated biological motifs?
  • How do generated sequence activity, novelty, and MMD vary across different divergence levels? Could similar divergence control be achieved by adjusting standard sampling parameters, such as temperature, sampling steps/schedule, or noise scale?
Reviewer_j45cOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Mutation Flows combines block insertions, planner-denoiser decomposition, and divergence conditioning for variable-length biological sequence generation. This paper takes antimicrobial peptide (AMP) as an application for mutation flows. Computational experiments on antimicrobial peptide (AMP) assess trajectory quality, few-step sampling, and predictor-guided optimization. Comparisons with edit-based generators and AMP design methods suggest improvements in distributional fidelity and selected activity prediction metrics.

Critical strengths

Mutation Flows architecture supports the reuse of pretrained denoisers, as demonstrated by the frozen ESM-2 experiment with promising results. Evaluation of intermediate trajectories and few-step sampling is relevant to predictor-guided optimization. The paper identifies Hyformer as the optimization objective and evaluates generated peptides with amPEPpy and APEX, which are not used for guidance.

Critical weaknesses

1): Validation is restricted to AMP design, with training sequences of 10 to 50 residues. Benefits for DNA, RNA, and longer protein sequences remain untested, which can not be trustworthy for Mutation Flows architecture in all biological sequence.

2): HydrAMP includes AMP and low-MIC probability filters in its official implementation, but the filtering status of the evaluated pool is unclear. Mutation Flows uses 200,000 nominal candidate-scoring evaluations for 1,000 final peptides. Equal final sample counts therefore leave search-budget comparability unresolved.

3): Hyformer guides optimization and also evaluates activity, limiting the independence of that metric. APEX-predicted MIC improves from 106.9 to 80.8 µM but remains higher than 56.2 µM for Hyformer. These results support metric-specific gains in predicted activity.

4): Novelty is based on exact matches against only 2,000 reference peptides, providing limited evidence of novelty relative to the complete training set.

Questions for authors

1): Mutation Flows is stated as general methods in biological sequence task but it only shows the protein task on antimicrobial peptide (AMP). Please add a DNA or RNA task with functional evaluation and appropriate baselines, or narrow the applicability claims. DNA-Diffusion (https://github.com/pinellolab/DNA-Diffusion), which has published on Nature Genetics, 2026, provides relevant context for regulatory DNA generation.

2): Please check HydrAMP (https://github.com/szczurek-lab/hydramp) official document, generation mode, decoding budget, and filtering criteria. Clarify whether the AMP and low-MIC probability filters in the official implementation were applied. Compare methods under matched proposal or predictor-query budgets while preserving the native selection procedures of each method.

3): Please clarify which metrics support the state-of-the-art claim and distinguish the Hyformer optimization objective from evaluations using amPEPpy and APEX. Explain how the APEX-predicted MIC of 80.8 µM, compared with 56.2 µM for the Hyformer baseline, affects the scope of the claimed activity improvements.

4): Please recompute novelty against the complete training set and compare it with the estimate based on the 2,000-peptide reference subset. Report nearest-neighbor similarity and include source-retention or corpus-resampling baselines to assess whether the novelty metric distinguishes generated sequences from training-set reuse.

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

Mutation Flows is a variable-length discrete flow-matching framework with three stated contributions: block insertions whose lengths follow a zero-truncated Poisson and resolve in one transition; an exact planner–denoiser decomposition that splits the objective into disjoint-parameter losses, letting a pretrained model serve as a drop-in denoiser; and divergence conditioning for coarse control over distance from the source sequence. Training uses 871,918 antimicrobial peptides of 10–50 residues. Evaluation covers trajectory MMD in ESM-C embedding space, few-step generation, a frozen ESM-2 denoiser, and predictor-guided activity optimization.

Critical strengths

The planner–denoiser split is well motivated. Separating where to edit and how long an insertion should be from which residues to actually place is the right factorization for variable-length generation, and it pays off concretely in the frozen ESM-2 experiment. Evaluating intermediate trajectory states, rather than only the endpoints, also fits the predictor-guided search setting the method is aimed at.

Critical weaknesses
  1. Table 1 does not support one of the paper's claims. Appendix B presents the in-house Edit Flows baseline as isolating "the contribution of block insertions", but under [U] that single-token baseline beats Single Model at every budget (1.58 / 1.18 / 0.88 / 0.70 / 0.68 / 0.60 against 1.71 / 1.40 / 1.07 / 0.94 / 0.80 / 0.80). So on the comparison the paper itself designates for its headline contribution, adding block insertions makes things worse.

  2. The few-step result rests on one metric in one regime. There is no second quality measure at matched budgets, such as pLDDT. For the sequence length, Appendix B restricts the data to 10–50 residues, which is relative short. Please add a second quality measure at matched NFE, and show few-step behaviour on longer sequence generation.

  3. Edit-aware diffusion for protein sequences should be discussed. The evolutionary modeling through local edits, and the factorization that predicts edit positions before token content, both have direct precedent in [1]. Its own stated limitation, that multi-residue insertions have to be composed across successive steps, which is exactly the gap block insertions fill here. I suggest the authors provide a comprehensive discussion about [1] and proposed method, otherwise it is hard to judge how novel the motivation and the planner–denoiser split really are.

[1] Towards A Generative Protein Evolution Machine with DPLM-Evo. ICML 2026

Questions for authors

Please see the above weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.