acceptodds
Under review as a conference paper at ICLR 2027

ShrinkFlow: Learning to Shrink Proteins from Protein Language Model Feedback

Abstract

Protein shrinking seeks shorter variants of existing proteins, yet experimentally validated source–shortened pairs are scarce and substantial shortening may require coordinated deletions across multiple regions. We introduce SHRINKFLOW, a framework for learning reusable multi-region shrinking policies from protein language model (PLM) feedback. A frozen PLM evaluates complete shortened candidates relative to their source, and the resulting plan preferences are converted into history–action supervision for a conditional discrete interval flow. A feasible action space enforces exact deletion budgets while preserving the identity and order of retained residues. Across general proteins and protein complexes, SHRINKFLOW achieves improved predicted structural and interface preservation relative to recent learning-based shrinking methods. Under shared functional-site constraints, it also preserves the local structural environment around protected catalytic regions more effectively. Controlled studies show that complete-outcome evaluation improves candidate selection over WT-only scoring and that the released policy achieves higher first-output structural quality than a capability-stage reference. Together, these results support reusable, exact-budget protein shrinking from sequence-model feedback without experimentally paired shortened targets.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein design, protein language models, protein shrinking, discrete flow matching

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_HbruOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces ShrinkFlow, a framework for deletion-only protein shrinking under an exact residue-removal budget. A frozen protein language model evaluates complete shortened sequences through global, pooled retained-residue, and residue-wise representational consistency. These evaluations are converted into proposal-corrected, plan-weighted history–action supervision. A conditional discrete interval-flow model then generates deletion intervals sequentially, with a right-to-left serialization and a feasibility-constrained action space that preserves retained-residue identity and order, supports protected positions, and guarantees completion for initially feasible requests.The evaluation covers general proteins, protein complexes, site-protected enzymes, higher-removal stress tests, and eight application targets. The paper reports improved predicted structural preservation over SCISOR and Raygun, and includes controls examining candidate scoring, proposal strategies, deletion morphology, and training supervision. The evidence is computational: activity, stability, and binding preservation are not experimentally established.

Critical strengths
  1. A coherent formulation method that decouples hard constraints from learning preferences.
    Interval representation, unique right-to-left serialization, and actionable support provide a clear implementation pathway for precisely controlling deletion budgets while preserving protected residues. Learning via full-plan preferences eliminates the need for action-level quality labels; meanwhile, one-time encoding of wild-type sequences enables an efficient route for distributing computational costs in result evaluation. This formulation is highly practical for constrained design problems: its effectiveness does not rely on the accuracy of quality models.

  2. Achieves significant improvements over evaluated learning-based baselines, validated through meaningful controlled experiments.
    The reported structural gains are notably superior compared to SCISOR and Raygun. For instance, when removing 40% of residues on the General-240 dataset, ShrinkFlow achieves a full WT TM score of 0.402, compared to 0.294 for SCISOR and 0.236 for Raygun-500. Under shared functional site protection conditions, ShrinkFlow shows a mean local Cα improvement of 0.2176 over SCISOR across budgets, with a 95% confidence interval of [0.1346, 0.3102], outperforming SCISOR in 19 out of 20 proteins. This protected-site experiment is particularly valuable as it distinguishes between retaining specific residues and maintaining their surrounding structural environments. These results deliver tangible and effective improvements across reported predictive structure endpoints.

Critical weaknesses
  1. Limited evidence that ShrinkFlow provides substantial advantages over strong simple baselines.Although ShrinkFlow outperforms existing learning-based methods, strong heuristic baselines remain competitive. Terminal truncation achieves better structural preservation when feasible, and under protected constraints the advantage over simple sampling strategies becomes small. This makes the practical benefit of the learned policy less clear.
  2. The contribution of PLM feedback distillation is not fully isolated.Matched supervision experiments show only modest improvements of Full over NoReward/Shuffled variants, suggesting that the main gains may come from the interval-based formulation rather than the proposed feedback learning itself. More analysis is needed to quantify the specific contribution of PLM feedback.
Questions for authors
  1. Could the author please explain under what circumstances ShrinkFlow has a significant advantage over terminal truncation or simple sampling strategies? Particularly, is this method significantly superior to these benchmark methods when internal deletion is required?
  2. How much of the final performance improvement comes from the PLM feedback supervision itself? Could the author provide more robust ablation experiments to distinguish the impact of complete result feedback from the interval generation architecture?
4 comments
AuthorsOpenReview ↗

For Question 1

We thank the reviewer for this question.

For some proteins, the termini contain relatively long extensions or flexible and disordered tails. When these regions do not serve essential functions, directly removing terminal segments can be a simple and effective way to shorten the protein. This strategy also has practical precedents in protein engineering. For example, Li et al. (2024) constructed N-terminally truncated TdT variants through PCR amplification and cloning. The truncated variants retained catalytic activity while exhibiting improved protein expression.

From an algorithmic perspective, ShrinkFlow's interval representation can directly accommodate N- and C-terminal deletions. Its generation strategy could also be adapted to first allocate part of the deletion budget to terminal regions and then use the remaining budget to generate internal deletion intervals, while strictly satisfying the prescribed total deletion budget. However, as in SCISOR and Raygun, we start from the original protein sequence without treating terminal truncation as a separate preprocessing step. Our goal is to investigate a more general protein-shrinking approach that handles terminal, internal, and multi-region deletions within a unified interval-generation framework, providing a potential direction for future research on general-purpose protein shrinking.

Moreover, protein termini are not always dispensable. N-termini may contain signal peptides involved in protein targeting and translocation (Ono, 2024). Therefore, the suitability of terminal regions for deletion cannot be determined solely by their sequence positions. For a more comprehensive experimental comparison, we have also included a terminal-only deletion baseline in Appendix F.3. In the functional-site protection experiments, terminal-only deletion produces feasible designs for only 28/36, 23/36, and 20/36 proteins at deletion ratios of 20%, 30%, and 40%, respectively, whereas ShrinkFlow satisfies the corresponding constraints for all 36 proteins at all three deletion ratios.

Regarding simple sampling, we provide two complementary comparisons. First, in Section 4.4 and Appendix F.2, we compare ShrinkFlow with a non-learning reference distribution (πL,B\pi_{L,B}), using the same number of candidates (K=32K=32) and the same complete-outcome PLM scorer. On 24 development proteins, ShrinkFlow achieves Full-WT TM scores of 0.619, 0.555, and 0.461 at deletion ratios of 20%, 30%, and 40%, respectively, outperforming the reference distribution by 0.031, 0.057, and 0.072. Native-contact recovery is also improved by 0.8, 9.9, and 15.0 percentage points, respectively.

Second, in the 36-protein functional-site protection experiments in Appendix F.4, we further compare ShrinkFlow with Protected Reference Sampling. When using the first generated output without candidate reranking, ShrinkFlow improves Local Cα and Native-contact recovery by 0.243 and 0.167, respectively. The gap becomes smaller when comparing against the stronger Sequential-uniform sampling baseline under the protected setting with (K=32K=32) candidate selection. This is because functional-site protection constraints, minimum interval lengths, and the exact deletion budget jointly impose substantial restrictions on the feasible action space F(h)\mathcal{F}(h). Sequential-uniform sampling also operates within our legal interval representation and uses the same complete-outcome PLM scorer to select among 32 candidates.

Taken together, these results show that when protected functional sites strongly constrain the feasible interval space and multi-candidate reranking is available, our legal interval action space combined with the complete-outcome PLM scorer already provides strong structural preservation. In contrast, ShrinkFlow's learned interval-selection policy offers clear advantages in general multi-region deletion tasks with larger search spaces, internal deletion settings, and first-output generation without candidate reranking (K=1K=1). Its central contribution is to learn interval-selection strategies from protein representations and deletion histories, enabling more informed multi-region deletion plans under exact deletion budgets and functional constraints.

References

Li, A.-N., Shi, K., Zeng, B.-B., Xu, J.-H., & Yu, H.-L. (2024). Enhancing the expression of terminal deoxynucleotidyl transferases by N-terminal truncation. Biotechnology Journal, 19(9), e2400226. https://doi.org/10.1002/biot.202400226

Ono, K. (2024). Signal Peptides and Their Fragments in Post-Translation: Novel Insights of Signal Peptides. International Journal of Molecular Sciences, 25(24), 13534. https://doi.org/10.3390/ijms252413534

AuthorsOpenReview ↗

For Question 2 (Part 1)

We thank the reviewer for raising this question. We have further investigated the role of PLM feedback in ShrinkFlow and added matched ablation experiments to examine the contributions of learned interval generation, complete-outcome quality supervision, and inference-time candidate scoring to the final performance. The new experimental results are reported in Table 24 of Appendix I in the revised manuscript.

The Role of PLM Feedback in ShrinkFlow

First, PLM feedback plays two distinct roles in ShrinkFlow: (1) providing complete-outcome quality supervision during generator training, and (2) evaluating and selecting the final deletion plan during multi-candidate inference.

During multi-candidate inference, the generator produces (K) deletion plans. We use a frozen PLM to jointly evaluate the global representation, retained-region representation, and residue-level representation consistency of each complete shortened sequence, and select the highest-scoring plan. We compare two scoring strategies on exactly the same candidate set. WT-only scoring uses the PLM representations of the original protein to estimate plan quality based on the deletion locations and the original representations of retained residues. In contrast, Complete-outcome scoring re-encodes the complete shortened sequence with the PLM and evaluates the representational changes between the original and shortened proteins.

Table 4. Comparison of the two complete-outcome scoring strategies (General-240, 30% deletion)

Scoring strategyFull-WT TM ↑Cα-lDDT ↑Native-contact recovery (%) ↑
WT-only0.4760.54535.99
Complete-outcome0.5760.75549.79

At a 30% deletion ratio on General-240, Complete-outcome scoring improves Full-WT TM from 0.476 to 0.576, Cα-lDDT from 0.545 to 0.755, and Native-contact recovery from 35.99% to 49.79%. Since both scoring strategies operate on exactly the same candidate set, these results indicate that evaluating the complete shortened sequence provides a more effective way to identify deletion plans with better structural preservation.

During generator training, we also use a frozen PLM to evaluate complete deletion plans and adjust their training weights according to their quality. Table 19 evaluates the contribution of this training-stage quality feedback, rather than removing all PLM-related mechanisms. Full adjusts training weights based on the quality of complete shortened sequences. NoReward retains the same generator architecture, original-protein PLM representations, and sampling-distribution correction, but removes the influence of complete-outcome quality scores on the training weights. WT-only replaces complete-outcome scoring with scoring based solely on the original protein representations, while Shuffled randomly permutes the correspondence between quality scores and deletion plans. All four settings use the same capability-pretraining initialization, training data, and training configuration, with three matched random seeds.

On General-58, we directly evaluate the eight unfiltered candidates generated by each training seed, using Native-contact recovery as the primary metric. Full consistently outperforms WT-only across all three deletion ratios and also shows positive gains over NoReward and Shuffled. For example, at a 30% deletion ratio, Full improves Native-contact recovery by approximately 0.6 and 0.4 percentage points over WT-only and NoReward, respectively. These results indicate that complete-outcome PLM feedback improves generator training, although the additional structural gains are relatively small. Table 19 measures the contribution of quality-weighted supervision during training and does not include PLM-based evaluation and selection during multi-candidate inference.

In Appendix F.4, we compare ShrinkFlow against a control that preserves the same deletion-interval configuration but randomly changes the deletion locations. The results show that interval configuration plays an important role in structural preservation, while the learned policy provides additional improvements in some structural metrics even when the interval configuration is matched. Appendix F.2 further compares learned Flow generation with sampling from a legal reference distribution, using the same number of candidates and the same PLM scoring function. At a 30% deletion ratio with (K=32), Flow improves Full-WT TM by 0.057, demonstrating the additional value of learned candidate generation over reference sampling.

AuthorsOpenReview ↗

For Question 2 (Part 2)

2. Additional Ablation Experiments

To further investigate the contributions of different design choices to final performance, we conducted six additional matched ablation experiments on General-58. We considered three candidate-generation strategies: Reference πL,B\pi_{L,B}, NoReward, and Full. For each strategy, we generated a candidate set and independently selected the final plan using either WT-only or Complete-outcome scoring.

Reference πL,B\pi_{L,B} is a legal reference distribution that does not rely on a learned generator. It first samples uniformly over the possible numbers of deletion intervals and then uniformly samples a complete deletion plan from all feasible plans with the selected number of intervals, subject to the exact deletion budget and minimum interval-length constraints. Thus, Reference πL,B\pi_{L,B} preserves the legal interval structure required by the protein-shrinking task but does not learn protein-specific deletion preferences from PLM representations. NoReward uses a learned interval generator without complete-outcome quality feedback during training, whereas Full uses a generator trained with complete-outcome quality feedback.

Six additional ablation experiments: 30% deletion ratio

Candidate generationScoring strategyFull-WT TM ↑Cα-lDDT ↑Contact ↑
Reference πL,B\pi_{L,B}WT-only0.34870.3040.2097
Reference πL,B\pi_{L,B}Complete-outcome0.46420.56650.3869
NoRewardWT-only0.46040.5870.402
NoRewardComplete-outcome0.50890.67830.4618
FullWT-only0.46440.59630.4059
FullComplete-outcome0.50990.68130.4665

First, comparing Reference πL,B\pi_{L,B} with NoReward measures the contribution of learned interval generation over legal reference sampling. At a 30% deletion ratio, under WT-only selection, NoReward improves Full-WT TM, Cα-lDDT, and Contact by 0.1117, 0.2830, and 0.1923, respectively. Under Complete-outcome selection, the corresponding improvements are 0.0447, 0.1118, and 0.0749. These results show that learned interval generation can produce candidates with better structural preservation even without complete-outcome quality feedback during training.

Second, comparing Full with NoReward measures the additional contribution of complete-outcome quality feedback during training. Under WT-only selection, Full improves the three metrics by 0.0040, 0.0093, and 0.0039, respectively. Under Complete-outcome selection, the corresponding improvements are 0.0010, 0.0030, and 0.0047. These results indicate that the additional structural gains from training-stage quality feedback are relatively small.

Third, comparing Complete-outcome with WT-only scoring on identical candidate pools isolates the contribution of inference-time PLM evaluation. At 30% deletion, Complete-outcome consistently improves all three structural metrics across Reference, NoReward, and Full. For Full, Full-WT TM increases from 0.4644 to 0.5099, demonstrating the value of complete-outcome scoring for candidate selection.

3. Discussion and Broader Implications

We suggest two possible reasons for the modest gains from complete-outcome quality weighting during generator training.

First, PLM feedback transfer to the generator could be improved. The PLM evaluates complete plans containing multiple deletion intervals, whereas the generator learns to select individual intervals. Plan quality determines the sampling weight, after which each plan is decomposed into sequential history–action examples. Consequently, all actions within a plan share the same quality preference, although their contributions may differ. Some plan-level quality information may therefore be lost during action-level supervision. Our experiments show that PLM feedback changes training-plan weights, but this has not yet translated into comparable improvements in generation quality.

Second, multi-region protein deletion involves a large combinatorial search space, while PLM evaluation of complete plans is computationally expensive. This limits the coverage and quality diversity of our candidate sets. Since quality weighting redistributes training probabilities among existing candidates, their coverage and quality distribution may limit feedback effectiveness. Expanding high-quality candidate coverage and developing finer-grained mechanisms for transferring plan-level preferences to action-level supervision are promising future directions.

Finally, we emphasize that ShrinkFlow aims to explore a general-purpose protein-shrinking framework that converts complete-outcome quality evaluation into learnable generation supervision and combines it with constrained interval generation to achieve multi-region protein shrinking under exact deletion budgets. We hope this framework provides an extensible research paradigm for incorporating structural information, functional evaluations, and eventually experimental measurement feedback into future protein-shrinking methods.

The additional comparisons with terminal truncation and reference sampling clarify the advantages and limitations of ShrinkFlow. The new ablation experiments also help distinguish the contributions of interval generation, PLM feedback, and inference-time scoring. My main concerns have been sufficiently addressed, and I have raised my score accordingly. I would like to state that AI was not used for the review process.

Reviewer_8NDpOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work introduces ShrinkFlow, a framework designed to learn multi-region shrinking policies for protein sequences using feedback from a frozen protein language model. The approach evaluates a complete deletion plan (a set of maximal, separated intervals summing to a given budget) along three cosine similarity axes against the wild type protein sequence, the cosine similarities are standardized into an overall plan score, and these scored deletion plans are converted into history-action supervision pairs. A conditional discrete interval flow is trained on these history-action pairs over a graph of feasible intervals, in order to ensure exact-budget termination. Evaluations across general proteins (General-240), protein complexes (Complex-150), and site-protected enzymes show that ShrinkFlow achieves higher predicted structural and interface preservation compared to existing methods such as SCISOR and Raygun.

Critical strengths
  1. Feasibility-constrained action space: The feasibility-constrained action space guarantees by construction that every permitted action preserves a valid completion, giving exact-budget generation. This is not done through rejection sampling or post-hoc revision of actions, but is rather constructed this way to preserve valid completions.
  2. Strong reporting of results and comparisons: The authors present many experiments in the main text and Appendix faithfully, including in cases where methods such as terminal truncation perform better relative to ShrinkFlow (Table 21). This makes the results more believable, and the experimental comparisons of ShrinkFlow relative to other baselines are thorough.
Critical weaknesses
  1. The results in the work are based completely on protein sequence model feedback, which limits conclusions made in this paper to predicted structural preservation, rather than true structural preservation. Without direct experimental measurement of actual protein activity or stability, there is a ceiling on claims which can be made based on these experiments, as acknowledged by the authors in their Conclusion.
  2. In several cases, terminal truncation achieved higher whole-chain structural scores than ShrinkFlow (Appendix F.3), indicating that ShrinkFlow constraints can hinder performance.
Questions for authors
  1. Can the sampling proposal be adjusted to allow for terminal truncations which would be beneficial for overall performance, rather than restricting to exact-budget deletions?
2 comments

We thank the reviewer for raising this important point.

First, prior computational protein-shrinking methods, including SCISOR and Raygun, do not introduce terminal truncation as a separate preprocessing step. We follow the same general setting. We believe this reflects an important biological consideration: protein termini are not inherently dispensable, and their suitability for deletion depends on the specific protein rather than their sequence positions alone. For example, N-terminal signal peptides mediate protein targeting and translocation (Ono, 2024), while N-terminal modifications can regulate protein stability, targeting, and molecular interactions (Øye et al., 2025). Similarly, the C-termini of some proteins contain PDZ-binding motifs that mediate protein–protein interactions (Zambo et al., 2024).

Importantly, our exact deletion budget does not preclude terminal truncation. Instead, it allows terminal and internal deletions to be considered jointly. From an algorithmic perspective, a two-stage strategy is also entirely feasible: first, allocate part of the deletion budget to the N- and/or C-termini, and then use the remaining budget to generate internal deletion intervals, while strictly satisfying the prescribed total deletion budget. There is therefore no algorithmic obstacle to incorporating terminal truncation.

Our goal, however, is to investigate a more general protein-shrinking problem: given an exact deletion budget, how can we identify suitable deletions without imposing a predefined priority between terminal and internal regions? We therefore handle terminal, internal, and multi-region deletions within a unified interval-generation framework, allowing the model to determine where deletions should occur based on the protein itself. We hope this provides a useful direction for studying general-purpose protein shrinking.

Meanwhile, for a more complete comparison, we have included terminal-only deletion results in Appendix F.3 and added the corresponding structural metrics for both methods in the revised Table 21.

References

Ono, K. (2024). Signal Peptides and Their Fragments in Post-Translation: Novel Insights of Signal Peptides. International Journal of Molecular Sciences, 25(24), 13534. https://doi.org/10.3390/ijms252413534

Øye, H., Lundekvam, M., Caiella, A., Hellesvik, M., & Arnesen, T. (2025). Protein N-terminal modifications: molecular machineries and biological implications. Trends in Biochemical Sciences, 50(4), 290–310. https://doi.org/10.1016/j.tibs.2024.12.012

Zambo, B., Gogl, G., Morlet, B., Eberling, P., Negroni, L., Moine, H., & Travé, G. (2024). Comparative analysis of PDZ-binding motifs in the diacylglycerol kinase family. FEBS Journal, 291(4), 690–704. https://doi.org/10.1111/febs.16994

I appreciate the authors response, which addresses my main concerns and questions. I choose to retain my score.

Reviewer_rmEaOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3(+1)
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces ShrinkFlow for protein shrinking, which turns protein language model (PLM) evaluations of complete shortened sequences as supervision to train a flow matching model-based multi-region deletion policy. Experiments show that SHRINKFLOW achieves higher predicted structural preservation than the evaluated learning-based shrinking methods, including improved interface-contact recovery and better preservation of local catalytic structure under shared constraints across general proteins, protein complexes, and site-protected enzymes.

Critical strengths
  1. The idea of obtaining plan-weighted history–action supervision from PLM evaluations and uses it to train a conditional discrete interval flow over feasible actions is interesting and reasonable.
  2. ShrinkFlow presents promising improvement on structural preservation of shortened sequences.
Critical weaknesses

The technical part of the paper is not easy to follow, especially the design and training of conditional discrete interval flow. It needs significant rewriting. Fig. 1’s symbols are not consistent with the technical part. Some technical details in the appendix may need to be moved to the main text for better understanding, e.g., the definition of ss.

Other points:

  1. PLM’s evaluation only reflects the feedback in terms of sequence. It would also be interesting to consider the feedback in terms of structure.
Questions for authors

The paper requires significant improvement in technical writing.

3 comments

We thank the reviewer for the helpful suggestions. We have revised Figure 1 and the Method section to provide a clearer explanation of the design and training of Conditional Discrete Interval Flow.

Specifically, we have revised Figure 1 to improve the correspondence between the notation in the figure and the main text, and to better illustrate the relationships among complete-plan scoring, weighted history–action supervision, and conditional interval-flow training. In Section 3.1, we have clarified the key notation and the definition of the feasible action set. In Section 3.3, we have added an explicit formulation of the sequential proposal probability and further explained how weighted history–action supervision is constructed from complete-plan scores. In Section 3.4, we have more clearly distinguished the inner interval-flow process from the outer sequential deletion process, and expanded the explanations of the interval graph, conditional transition-rate network, and path-based training procedure. In Section 3.5, we have provided additional details on the training setup and inference-time generation.

More detailed algorithmic implementations and derivations are retained in the Appendix. The revisions are highlighted in orange in the revised manuscript.

Thanks for the careful revision of the paper.

I have gone through the revised manuscript, and the technical part is now much clearer. I would suggest adding an algorithm to further improve clarity.

The idea of obtaining plan-weighted history–action supervision from PLM evaluations and using it to train a conditional discrete interval flow over feasible actions is quite impressive. So I recommend acceptance and increase my rating accordingly.

We sincerely thank the reviewer for taking the time to carefully examine our revised manuscript and for the encouraging feedback. We are glad that the revisions have improved the clarity of the technical presentation. We also appreciate the suggestion and would add a concise algorithm summarizing the training and inference procedures to the camera-ready version if the paper were to be accepted.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.