- 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
-
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.
-
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.
-
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?
-
Does the network receive time separately? Appendix B gives only , the requested divergence multiplied by the remaining schedule fraction , whereas Proposition 3 conditions on both time and divergence. Different pairs share the same . Please make sure the implementation and the proof agree.
-
Since source sequences already follow the target distribution, please include unchanged-source and corpus-resampling controls. Report actual source-to-output divergence 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.
-
Algorithm 2 uses no-edit probability , where is the step size and is the total edit rate at position . 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.
-
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.
I look forward to the response from the authors and am willing to revise my score with clarifications to my questions.