acceptodds
Under review as a conference paper at ICLR 2027

LaCPlow: Geometry-Preserving Partially Latent Flow Matching for Cyclic Peptide Generation

Abstract

Cyclic peptides share much of their local stereochemistry with linear peptides, but cyclization alters their covalent topology and reshapes their global conformations through sparse, nonlocal chemical constraints. Satisfying these constraints depends on the relative placement and orientation of linked residues. We propose LaCPlow, a target-conditioned partially latent flow model that explicitly transports residue frames while encoding sequence and residue-local atom coordinates in a continuous latent representation. A representation-aligned joint vector field couples these components by predicting rotational, translational, and latent velocities. Because chemistry-specific specialization depends on the capabilities of its substrate, we first evaluate LaCPlow-Base without cyclization-specific supervision to assess its geometric modeling capacity. We then introduce LaCPlow-Cyc, which incorporates cyclization-specific supervision. LaCPlow-Cyc improves bond-forming geometry while largely preserving the stereochemical quality of the base model.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
cyclic peptide design, flow matching, partially latent generative models, geometric deep learning, SE(3)-equivariant modeling

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_HQq7OpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work focuses on target-conditioned cyclic peptide generation and proposes a geometry-aware partially latent flow matching framework (LaCPlow). Specifically, the method explicitly models the rigid frames of amino acid residues, including their spatial positions and rotational orientations, as geometric variables, while employing a variational autoencoder (VAE) to encode amino acid identities and residue-local atomic coordinates into continuous latent variables. Building on this representation, the method constructs a geometrically equivariant vector field that jointly models residue rotations, translations, and latent variables, enabling the joint generation of amino acid sequences and all-atom structures conditioned on a target protein through flow matching. Furthermore, the work introduces a geometry-aligned attention network to capture spatial interactions both within the cyclic peptide and between the peptide and its target protein. It additionally incorporates cyclization-specific supervision and ring-closure geometric constraints to improve the cyclization success rate and stereochemical plausibility of the generated cyclic peptides.

Critical strengths
  1. The proposed modeling paradigm demonstrates a certain degree of novelty and practical effectiveness compared to prior work. Specifically, the method explicitly models the flow of residue translations and rotations while encoding residue identities and local atomic structures into latent variables, enabling their joint generation.

  2. The experimental evaluation assesses the quality of generated structures before energy minimization or other relaxation procedures. This allows for a more precise assessment of the model’s intrinsic generative capabilities and the additional contributions of post-processing.

  3. The algorithmic design and experimental setup effectively distinguish between the geometric generation capabilities of the base model and the specialized generation capabilities tailored to specific chemical constraints.

  4. The ablation studies provide strong empirical evidence for the effectiveness of several key design choices. Under the same $(R,x,z) $ representation, the proposed geometric network outperforms the PairBiased Transformer. Furthermore, incorporating cyclization information through latent-variable-based modeling achieves a higher ring-closure success rate than treating cyclization information as an additional explicit flow state.

Critical weaknesses
  1. The methodological novelty of LaCPlow is relatively incremental. Its main contribution lies in the targeted integration of explicit residue rigid frames, partially latent flow matching, and a representation-aligned geometric network, rather than the introduction of a fundamentally new generative modeling paradigm.

  2. The independent contribution of explicitly modeling residue rotations R has not been rigorously validated. A controlled comparison between (x,z)(x,z) and (R,x,z)(R,x,z) under identical training data, model capacity, and training configurations is needed to isolate the benefits of introducing explicit rotational variables.

  3. The contributions of individual components within the proposed geometric network have not been sufficiently investigated. Although Table 4 demonstrates that the complete geometric network outperforms the PairBiased Transformer, it does not adequately examine the individual contributions of its key architectural components.

  4. The necessity and effectiveness of independently sampled time variables remain insufficiently validated. The proposed joint flow matching framework employs two independently sampled time variables for geometry and latent representations. However, the experiments do not directly compare shared time scheduling (tT=tz)(t_T=t_z) with independent time scheduling (tT≠tz)(t_T\neq t_z). Consequently, the current results do not establish whether independent time scheduling offers advantages over shared scheduling, nor do they clarify its specific contributions to generation quality, training stability, or the coupling between geometric and latent representations.

Questions for authors
  1. Sufficiency of existing methods. Can existing protein co-design models achieve performance comparable to LaCPlow after fine-tuning on cyclic peptide data? If so, what is the necessity and additional value of introducing the proposed architecture?

  2. Necessity of joint atomic geometry modeling. Why is it necessary to jointly generate residue-local atomic geometry? What are the specific advantages of this approach over a two-stage pipeline that first generates the backbone and then performs all-atom reconstruction and side-chain completion?

Post-Rebuttal Updates: The rebuttal addresses most of my concerns and provides sufficiently interesting and convincing findings to justify raising my rating to 3. However, I am not an expert in this specific area, and I encourage the AC to take my limited expertise into account.

4 comments

Dear reviewer HQq7, we sincerely thank you for the constructive feedback.

[W1] The novelty of LaCPlow is incremental

The central scientific novelty of this work is to question whether existing peptide or protein generation architectures provide a sufficient foundation for cyclic peptide modeling, even when augmented with cyclization-specific components. LaCPlow-Base already achieves better cyclic peptide generation than models incorporating such components. This finding shows why the underlying generative architecture deserves attention alongside the mechanisms added to enforce cyclization.

Accurate cyclic peptide modeling requires an appropriate representation of residue orientations and a geometric network that can use this information effectively. Fully explicit approaches represent orientations directly, but can still struggle when orientation, conformation, and sequence are tightly coupled. Fully latent and other partially latent approaches handle residue orientations indirectly. LaCPlow explicitly transports residue frames, represents sequence and intra-residue geometry in latent variables, and uses a geometric network aligned with this representation. Its performance challenges the assumption that a linear-peptide generator can serve as a sufficient foundation, leaving additional cyclization-specific mechanisms to account for cyclic peptide geometry.

We also examine how post-generation refinement can blur the distinction between a generator’s intrinsic geometric quality and the corrections made by an energy minimizer. Evaluating structures before refinement allows us to assess what the generator itself has learned. This distinction also matters for optimization. Existing approaches, including APCyc, GeoCycler, and CP-Composer, pursue cyclic peptide optimization using underlying generative architectures whose ability to represent cyclic peptides remains limited. LaCPlow offers an alternative foundation for these workflows by combining an architecture designed to capture cyclic peptide geometry with a latent-space optimization interface. The contribution is therefore broader than an additional optimization strategy. It provides a generative substrate on which future cyclic peptide generation and optimization methods can be built, and shows why the representation and processing of cyclic peptide geometry are themselves important architectural choices.

[Q1] Comparison with an existing protein co-design model fine-tuned on cyclic peptide data

In the paper, we compared against protein and peptide co-design models retrained on cyclic peptide data. We additionally fine-tuned the official pretrained Proteina-Complexa model on cyclic peptide data using LoRA with the authors' official fine-tuning code. We evaluated the epoch-2 checkpoint, selected by validation loss.

modelPdbondP_{d_\mathrm{bond}}PCβP_{C_\beta}chiralityRama allowedRama Favored
test dataset10098.810097.389.8
LaCPlow-Base25.8‾\underline{25.8}95.5‾\underline{95.5}10095.3‾\underline{95.3}83.5‾\underline{83.5}
LaCPlow-Cyc6193.910094.583.5‾\underline{83.5}
proteina-complexa(from scratch)2096.399.9‾\underline{99.9}73.351.9
Proteina-Complexa(fine tuning)16.4891.0510099.6697.9
modelEnergydGVinaGRAVYdiversitynoveltyScRMSD(H2T)ScRMSD(C2C)
test dataset98.8-26.4-3.291.5-0.3963.182.70
LaCPlow-Base97.2‾\underline{97.2}−22‾\underline{-22}-4.1191.1‾\underline{91.1}0.20‾\underline{0.20}0.5413.523.63
LaCPlow-Cyc97.0-20.6−4.19‾\underline{-4.19}86.80.220.534‾\underline{0.534}3.483.6
proteina-complexa(from scratch)85.3-18.4-3.791.50.120.5082.52‾\underline{2.52}2.68
Proteina-Complexa(fine tuning)98.94-32.83-5.5979.480.0490.4252.512.75‾\underline{2.75}

The same Proteina-Complexa architecture trained from scratch on cyclic peptide data achieves better closure geometry. These results suggest that strong performance on properties shared with proteins does not necessarily transfer to cyclic peptide specific geometry.

[W2&W3] The independent contribution of representation and network components

Assessing the contribution of explicit residue orientations is important. Table 4 compares non-equivariant and equivariant architectures, two widely used approaches to protein generation. We thank the reviewer for pointing out that this broad comparison makes individual architectural contributions difficult to isolate. In LaCPlow, however, RR provides the coordinate frames used for geometric interactions and local reconstruction, rather than simply an additional input. Relative positions, relative rotations, and atom reconstruction depend on these frames. Removing RR therefore requires either estimating frames separately or replacing these operations with frame-independent alternatives. Our control uses the latter approach. A simple on/off comparison that leaves every other operation unchanged is consequently not well-defined.

PdbondP_{d_\mathrm{bond}}PCβP_{C_\beta}ChiralityRama allowedRama favoreddGenergy
A22.2‾\underline{22.2}94.1‾\underline{94.1}57.061.144.5-17.488.5
Bw/ equation (14)15.984.810090.7‾\underline{90.7}76.3‾\underline{76.3}−19.2‾\underline{-19.2}87.9
Cw/ equation (8)17.790.010089.475.2−19.2‾\underline{-19.2}92.1‾\underline{92.1}
D (=LaCPlow)w/ IPA25.895.510095.383.5-2297.2

Construction of control A.

Control A uses the generative state (x,z). Its peptide–peptide geometric features are reduced as follows: [Ri⊤(xj−xi);ρ(Ri⊤Rj);RBF(∣xj−xi∣)]⟶RBF(∣xj−xi∣)\left[R_i^\top(x_j-x_i);\rho(R_i^\top R_j); \mathrm{RBF}(|x_j-x_i|) \right] \quad\longrightarrow\quad \mathrm{RBF}(|x_j-x_i|).

Observed target frames remain available. PP and PT therefore use distance features only, TP retains directional features expressed in the target frame, and TT remains unchanged.

For peptide to target atom conditioning, local directional vectors are likewise replaced with distances: rima=Ri⊤(Xm,aP−xi)⟶dima=∣Xm,aP−xi∣r_{ima}=R_i^\top(X^P_{m,a}-x_i) \quad\longrightarrow\quad d_{ima}=|X^P_{m,a}-x_i|.

The network uses a PairBiased Transformer, which does not require explicit peptide frames. The translational readout and atomic decoder construct vectors as weighted sums of relative position vectors, using invariant scalar coefficients. This preserves SE(3)-equivariance without peptide frames.

In B, C, and D, the components of the network were evaluated.

[W4] comparison between shared training time and independent training time

We additionally trained a shared training time variant by sampling a single time t∼U(0,1)t\sim\mathcal U(0,1) and setting tT=tz=tt_T=t_z=t for each training example.

modelPdbondP_{d_\mathrm{bond}}PCβP_{C_\beta}chiralityRama allowedRama Favored
LaCPlow29.2097.0010095.6584.14
shared training time25.20‾\underline{25.20}95.63‾\underline{95.63}10095.05‾\underline{95.05}83.32‾\underline{83.32}

Both models were evaluated using LaCPlow’s default sampling schedule at inference (Appendix A.4), following the evaluation protocol of Table 8, with 100 samples for each of 128 CPCore validation pockets.

[Q2] Advantages of joint modeling over a two-stage pipeline

In a conventional two-stage pipeline, the backbone is generated first and held fixed during side-chain packing. This constrains subsequent side-chain placement and can limit the ability to satisfy closure geometry, leaving greater corrections to post refinement. For example, MuCO follows this two-stage pipeline, whereas LaCPlow jointly generates backbone and side-chain information. Table 3 shows that MuCO requires larger adjustments especially in closure distance to achieve the required cyclization geometry.

The table below directly compares LaCPlow and MuCO in Table 3.

modelBBSCBB bondSC bondBB torSC torδdbond\delta d_\mathrm{bond}δCβ\delta C_\beta
LaCPlow-Base2.152.5756.737.3‾\underline{37.3}47.5‾\underline{47.5}0.53‾\underline{0.53}0.87
LaCPlow-Cyc2.11‾\underline{2.11}2.52‾\underline{2.52}5.1‾\underline{5.1}6.8‾\underline{6.8}3746.70.260.92‾\underline{0.92}
MuCO1.261.5312.131.945.457.33.71.29

The table below compares overall structural quality.

modelPdbondP_{d_\mathrm{bond}}PCβP_{C_\beta}chiralityRama allowedRama Favored
LaCPlow-Base25.8‾\underline{25.8}95.510095.3‾\underline{95.3}83.5‾\underline{83.5}
LaCPlow-Cyc6193.9‾\underline{93.9}10094.583.5‾\underline{83.5}
MuCO2.878.750.3‾\underline{50.3}96.289.4
Reviewer_HQq7OpenReview ↗

Thanks for the Rebuttal

Thank you to the authors for the detailed rebuttal and additional experiments.

I find the consolidated evaluation for Q1 particularly interesting and convincing. Although the fine-tuned Proteína-Complexa produces more candidates passing the strict joint filtering criteria, LaCPlow-Cyc achieves broader target coverage, less concentration of valid candidates, and greater structural diversity across targets. These results strengthen the practical value of the proposed architecture.

I am therefore willing to raise my rating from 2 to 3.

My remaining concern is whether the Proteína-Complexa baseline has been sufficiently fine-tuned. Could full-parameter supervised fine-tuning, rather than LoRA, yield better results? Were alternative fine-tuning strategies explored?

I will maintain my original confidence score, as cyclic peptide generation is not my primary research area, and I encourage the Area Chair to take this into account.

Dear Reviewer HQq7,

We sincerely thank you for raising your rating, and for the follow-up question on full-parameter fine-tuning.

Following Proteina-Complexa’s original training setup, we kept the released autoencoder frozen and fine-tuned only the flow model in both variants.

We additionally applied full-parameter fine-tuning on the same CPCore training split, and evaluated the epoch-2 checkpoint, selected by validation loss, using the protocols in our Consolidated Evaluation. We report only the additional rows below to avoid repeating the full comparison tables.

Modelnumber of attemptsGeneration successF1F2F3F4F5F6F7F8F9R1R2R3R4R5R6
Proteina-Complexa (full-param tuning)8400840019051873807156109108108108105391.25032.451.41.3080.607

LaCPlow-Cyc achieves broader target coverage and a higher mean number of distinct passing structural clusters.

Modelgenerated samples/targetsamples retained after filtering/targetstructural diversity (clusters/target)sequence diversity (clusters/target)
Proteina-Complexa (full-param tuning)92.031.111.46

LaCPlow-Cyc produces more distinct structural and sequence clusters under the matched sampling-time budget, supporting its practical advantage.

Once again, we sincerely thank you for your time and for raising your rating, and we hope these additional results fully address your remaining concern. Thank you very much!

Reviewer_ZcxPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LaCPlow, a partially latent flow matching model for target-conditioned cyclic peptide generation. The method explicitly models residue rotations and translations while encoding amino acid sequences and local atomic structures in a latent representation. The authors further introduce cyclization-specific supervision to improve bond-forming geometry. Experiments show that LaCPlow achieves better closure rates and stereochemical quality than several baselines, and demonstrate the potential of latent space for structural optimization.

Critical strengths
  1. The paper provides comparisons with multiple baselines and evaluates several aspects of generated peptides, including closure rate, stereochemical quality, binding-related scores, and diversity. The experimental evaluation is relatively comprehensive.

  2. LaCPlow achieves strong results in cyclization geometry and structural quality. In particular, LaCPlow-Cyc reaches a closure rate of 61.0% in Table 2, significantly outperforming other methods while maintaining high chirality and Ramachandran quality.

Critical weaknesses
  1. The novelty of LaCPlow is somewhat limited. Its partially latent representation is similar to existing methods such as La-Proteina and Proteína-Complexa, and its geometric modeling mainly builds on existing equivariant architectures. In addition, general all-atom generative models such as PocketXMol can already handle cyclic peptide generation. Without a direct comparison with these methods, it is unclear what advantages the proposed architecture provides over existing approaches.

  2. LaCPlow shows relatively low structural diversity. In Table 12, LaCPlow-Cyc achieves a structural diversity of only 0.22, compared to 0.50 for PepFlow and 0.56 for APCyc. Although the authors argue that target-fixed diversity may be affected by binding pose variation, LaCPlow still shows lower diversity after structural alignment. The current experiments do not clearly show whether this limited diversity affects the model's ability to explore different high-quality cyclic peptides.

  3. The advantage of latent optimization is not sufficiently demonstrated. Section 4.6 shows that SPSA-based latent editing improves the interface score and Rosetta dG, but the experiments only compare it with random perturbations. These results show that the latent space can be optimized, but do not demonstrate its advantage over directly optimizing explicit geometry or other existing optimization methods.

Questions for authors
  1. Could the authors compare LaCPlow with general all-atom generative models such as PocketXMol on the same test set and evaluation metrics to better demonstrate the advantages of the proposed method?

  2. Could the authors compare the number of distinct high-quality cyclic peptides generated by different methods under the same sampling budget? For example, the authors could cluster samples that satisfy both closure and stereochemical constraints and compare the number of unique successful structural clusters. This would help clarify whether the lower structural diversity limits the practical design capability of LaCPlow.

  3. Could the authors compare latent editing with direct optimization of explicit residue frames (R,x) or other existing optimization methods under the same optimization budget? Comparing the final interface score, Rosetta dG, and structural quality would help demonstrate whether latent optimization provides an actual advantage.

4 comments

Dear reviewer ZcxP, we sincerely thank you for the constructive feedback.

[W1] The novelty of LaCPlow is limited

The central scientific novelty of this work is to question whether existing peptide or protein generation architectures provide a sufficient foundation for cyclic peptide modeling, even when augmented with cyclization-specific components. LaCPlow-Base already achieves better cyclic peptide generation than models incorporating such components. This finding shows why the underlying generative architecture deserves attention alongside the mechanisms added to enforce cyclization.

Accurate cyclic peptide modeling requires an appropriate representation of residue orientations and a geometric network that can use this information effectively. Fully explicit approaches represent orientations directly, but can still struggle when orientation, conformation, and sequence are tightly coupled. Fully latent and other partially latent approaches handle residue orientations indirectly. LaCPlow explicitly transports residue frames, represents sequence and intra-residue geometry in latent variables, and uses a geometric network aligned with this representation. Its performance challenges the assumption that a linear-peptide generator can serve as a sufficient foundation, leaving additional cyclization-specific mechanisms to account for cyclic peptide geometry.

We also examine how post-generation refinement can blur the distinction between a generator’s intrinsic geometric quality and the corrections made by an energy minimizer. Evaluating structures before refinement allows us to assess what the generator itself has learned. This distinction also matters for optimization. Existing approaches, including APCyc, GeoCycler, and CP-Composer, pursue cyclic peptide optimization using underlying generative architectures whose ability to represent cyclic peptides remains limited. LaCPlow offers an alternative foundation for these workflows by combining an architecture designed to capture cyclic peptide geometry with a latent-space optimization interface. The contribution is therefore broader than an additional optimization strategy. It provides a generative substrate on which future cyclic peptide generation and optimization methods can be built, and shows why the representation and processing of cyclic peptide geometry are themselves important architectural choices.

[Q1] comparison between existing all-atom models

We already compared LaCPlow with general all-atom generators, including Proteina-Complexa, and additionally evaluated PocketXMol using its officially released pretrained weights and H2T cyclic peptide generation protocol. This PocketXMol interface receives a predefined H2T backbone topology, including the closure bond, whereas LaCPlow generates target-conditioned cyclic peptides and determines their cyclization type without a supplied topology. PocketXMol’s geometric fidelity should therefore be interpreted conditional on this additional input.

modelgeneration successPdbondP_{d_\mathrm{bond}}PCβP_{C_\beta}chiralityRama allowedRama Favored
test dataset-10098.810097.389.8
PocketXMol7527/840084.187.946.8‾\underline{46.8}61.242.0‾\underline{42.0}
LaCPlow-Base8400/840025.895.510095.383.5
LaCPlow-Cyc8400/84006193.9‾\underline{93.9}10094.5‾\underline{94.5}83.5
modelEnergydGVinaGRAVYdiversitynoveltyScRMSD(H2T)
test dataset98.8-26.4-3.291.5-0.3963.18
PocketXMol89.6-28.1-4.7378.80.610.5793.52‾\underline{3.52}
LaCPlow-Base97.2−22‾\underline{-22}-4.1191.10.200.541‾\underline{0.541}3.52‾\underline{3.52}
LaCPlow-Cyc97.0‾\underline{97.0}-20.6−4.19‾\underline{-4.19}86.8‾\underline{86.8}0.22‾\underline{0.22}0.5343.48

The comparison with PocketXMol shows that achieving a high closure distance pass rate under a given topology does not necessarily guarantee the high stereochemistry. closure distance and stereochemistry should therefore be evaluated jointly. Meanwhile, LaCPlow-Cyc raises the closure distance pass rate while largely preserving the stereochemical quality of LaCPlow-Base. This indicates that additional cyclization-specific supervision can improve closure geometry without substantially degrading the stereochemistry of the base model.

[W2&Q2] Compare the diversity of valid samples between LaCPlow and the baselines under the same sampling budget.

We set the sampling budget to the time that the fastest model, APCyc, needs to generate 100 samples per target (88 seconds). Under this budget, we filtered the samples generated by each model using the criteria listed below and compared LaCPlow with the baseline models. The number of generated samples per target is rounded to the nearest integer.

Modelgenerated samples/targetsamples retained after filtering/targetstructural diversity (clusters/target)sequence diversity (clusters/target)
PepFlow281.281.271.28
DiffPepBuilder120.10.10.1
PepGLAD (raw)67000
PepGLAD (ideal.)670.040.040.04
Proteina-Complexa90.150.140.15
LaCPlow-Base449.66‾\underline{9.66}4.43‾\underline{4.43}9.61‾\underline{9.61}
APCyc (raw)100000
APCyc (c*)100000
CP-Composer57000
LaCPlow-Cyc4422.54722.09

Under this same budget and criteria, LaCPlow generated a larger number of distinct structural and sequence clusters compared to baselines.

[Filter]

dG<0 ∧ dbondd_{\mathrm{bond}} window ∧ CβC_\beta window ∧ Rama allowed ≥ 0.80 ∧ Rama favoured ≥ 0.70

[W3&Q3] Relative advantages of latent optimization in LaCPlow

We first clarify that our optimization experiment in Section 4.6 was not designed to establish superiority over direct optimization of explicit geometry or other existing optimization methods. The suggested comparison would be useful for assessing the relative benefits, but our experiment asks whether LaCPlow’s latent representation can serve as an optimization interface.

One motivation discussed in the Introduction is the representation-specific complexity of editing explicit generative states. In fully explicit approaches, variables belong to different spaces, so direct optimization requires choosing which variables to edit and defining suitable optimization methods and scales for each. Continuous latent variables offer a simpler interface for applying a range of optimization methods. This practical advantage has encouraged the continued use of architectures such as PepGLAD as generative substrates in approaches including CP-Composer, APCyc, CycDiff-DPO, and GeoCycler.

In a partially latent architecture, however, the presence of latent variables alone does not guarantee that editing them will also affect explicit geometry. We therefore first verified that interventions on the latent variables propagate through the joint flow to residue positions and orientations. We then showed that optimization improves the objective score relative to random perturbations. Improvements were also observed under an external scoring oracle.

These results support the use of LaCPlow’s latent representation as an optimization interface. We have not claimed its relative advantage over direct optimization of explicit generative states or other optimization methods. We will clarify this scope in the Limitations, noting that our experiments demonstrate the feasibility of latent optimization rather than its superiority over direct geometric optimization or other methods.

The additional experiments address some of my concerns, particularly regarding structural diversity. The comparison under the same sampling budget provides convincing evidence that LaCPlow can generate more distinct valid cyclic peptides than the baselines.

The PocketXMol comparison is useful, although the results are mixed and the input conditions differ. For latent optimization, the authors clarify that they only aim to demonstrate feasibility, not superiority over existing methods.

My main concern remains the limited methodological novelty. The rebuttal strengthens the experimental results but does not fully resolve this issue. I will therefore maintain my original rating of Weak Rejection.

Dear Reviewer ZcxP,

Thank you for your constructive feedback, which helped us strengthen the claims of our paper.

We understand that novelty is inherently subjective, and to address the remaining concern about it, we would like to cite the ICLR 2027 guidelines:

Does it contribute new knowledge and sufficient value to the community? Note, this does not necessarily require state-of-the-art results. Submissions bring value to the ICLR community when they convincingly demonstrate new, relevant, impactful knowledge (incl., empirical, theoretical, for practitioners, etc).

The perspective that properly handling the local residue frame through our methodology enables more precise modeling of cyclic peptides has been neither readily recognized nor explored. During the rebuttal period, we empirically validated our hypothesis precisely at the point where a representational/methodological gap exists. (Please note in particular our baseline models, which either use residue frames explicitly or approximate them.)

Among current cyclic peptide generative models, none has so far been able to generate samples of comparable quality to those of LaCPlow when trained from scratch on a cyclic peptide dataset (see Section 4.7 and Appendix D of the revised paper).

In conclusion, LaCPlow provides a new substrate architecture for a community where modeling the distribution of cyclic peptides remains an open challenge, and we believe it will accelerate cyclic peptide drug discovery and have a broader impact.

Reviewer_vk4BOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

LaCPlow generates pocket-conditioned cyclic peptides using flow matching over explicit residue frames together with latent sequence-related variables. A cyclization-aware variant introduces additional closure guidance and evaluates pre-minimization geometry separately from post-minimization structural and surrogate binding metrics. The evaluation spans 84 held-out targets with 100 samples per target and compares explicit, fully latent and partially latent peptide generators. The primary supported contribution is better geometric closure without giving up the favorable local geometry of partially latent generation. The key distinction is between raw generative geometry and physical/chemical usability. Flow matching over partially explicit residue frames and latent components can improve a pre-minimization closure proxy, but successful covalent cyclization additionally requires bond angles, stereochemistry and tolerable strain. The evaluation also filters some post-relaxation outcomes, making a common initial sample denominator crucial.

Critical strengths
  1. Ring closure is an essential distinction between cyclic and linear peptide generation; explicitly modeling the topology-relevant constraint is technically well motivated. The partially latent formulation also addresses local stereogeometry more directly than fully latent alternatives. The explicit geometric representation addresses a constraint that linear-peptide generation routinely ignores. It is also methodologically meaningful to guide closure during generation rather than use a minimizer to repair arbitrary open structures afterward. The formulation thus tests a biologically motivated inductive bias instead of relying solely on downstream postprocessing.

  2. The reported gains are substantial under the paper's pre-minimization geometric proxy. Table 1 shows LaCPlow-Base Pdbond closure-window success of 25.8%, and the cyclization-aware variant later reaches approximately 61% (Table 2), rather than merely obtaining a better relaxed energy after postprocessing. The 61.0% pre-minimization closure-window success directly measures the component the authors intend to improve. Comparators such as raw APCyc are much lower under this proxy, so the effect is not a small numerical fluctuation. This finding deserves credit even though terminal proximity alone does not establish chemical validity.

  3. The paper reports complementary chirality, Ramachandran, closure and diversity metrics, distinguishes raw versus minimized structures, and documents baseline training/checkpoint differences and filtering choices. This is valuable for assessing what each claimed improvement means. The reported chirality and Ramachandran metrics are relevant because a closure-only optimization could otherwise destroy local structural plausibility. Inclusion of diversity, energy/affinity surrogates, parameter size and sampling time prevents the claim from resting on one favorable column. The target-clustered split is also more credible than a random split of similar complexes.

Critical weaknesses
  1. Pdbond measures proximity within a distance window, not proof of covalent ring topology, chemically valid bond geometry, stereochemical consistency or acceptable strain. Roughly 39% of samples still fail even that primary pre-minimization closure proxy. The conclusion must be framed as improved geometry, not guaranteed chemically synthesizable cyclic peptides. A bond-distance window omits valence topology, terminal bond geometry, cis/trans state and ring strain. Even the window criterion fails for approximately 39% of raw candidates. An honest description is 'greater geometric closure probability' unless full stereochemical/topological validity is demonstrated. This is a substantive difference for users intending peptide synthesis or downstream docking.

  2. Accuracy, diversity and cost are in tension. Appendix Table 12 reports structural diversity around 0.22 for LaCPlow-Cyc versus 0.94 for CP-Composer under the authors' metric. The reported LaCPlow-Cyc configuration also has approximately 42.11M parameters and 2.01 s/sample versus 5.32M and 0.88 s/sample for APCyc. The paper should not infer an unequivocally superior generator from closure gains alone. The trade-off is large: Appendix Table 12 reports structural diversity 0.22 versus 0.94 for CP-Composer, and the cyclization variant is much larger/slower than APCyc. A model that frequently produces one narrow family of closed conformations may be less useful for target discovery than a diverse sampler with lower per-sample closure. Practical merit should therefore be judged on valid, diverse candidates per compute budget.

  3. Metrics use different denominators: closure/chirality describe all raw candidates, whereas certain affinity and diversity summaries are conditioned on minimization and a dG < 0 filter. This can obscure the joint rate of obtaining a closed, chemically valid, acceptable-affinity candidate. With multiple samples per target, uncertainty should be clustered by receptor/target rather than assuming 8,400 independent targets. Post-minimization affinity and novelty summaries conditioned on dG or other filters can select an easier subset, while closure is defined over all raw generations. Without a staged denominator table, a high conditional affinity score cannot be interpreted as high unconditional joint success. Multiple samples from each of 84 targets are also correlated, so target-clustered uncertainty is more appropriate than treating every generation as independent.

Questions for authors
  1. Please give a table starting from all N=84x100 raw generations, reporting the exact count passing each successively applied criterion: terminal distance, topology/bond-angle geometry, stereochemistry and strain, minimization, and any affinity filter. What is the unconditional joint yield per target, not only a mean conditioned on successful filtering?

  2. What is the frontier between closure yield, structural/sequence diversity and sampling seconds or FLOPs? In particular, can the authors fairly compare LaCPlow-Cyc with APCyc and CP-Composer under equal sampling budgets, noting any differences in baseline training/checkpoint or refinement conditions?

  3. Please report target-level means and confidence intervals, treating the 84 targets as the independent units rather than all 8,400 generated candidates. How many targets achieve at least one chemically acceptable closed candidate, and are improvements driven by a small subset?

5 comments

Dear reviewer vk4B, we sincerely thank you for the positive and constructive feedback.

[W1-1] Since PdbondP_{d_{\mathrm{bond}}} measures proximity within a distance window, it should not be described as guaranteeing chemically synthesizable cyclic peptides.

We thank the reviewer for this comment, and we fully agree that PdbondP_{d_{\mathrm{bond}}} should be interpreted as an indicator of a higher likelihood of geometric closure. We have shown that the PdbondP_{d_{\mathrm{bond}}} metric proposed in LaCPlow evaluates more selectively than the PCβP_{C_{\beta}} metric (see Appendix B.2), which has been treated as a success metric in the cyclic peptide community. Moreover, because it measures the atom pairs that actually form the bond, it corresponds more directly to closure chemistry. However, as stated in Appendix B.1, we would like to clearly acknowledge that even PdbondP_{d_{\mathrm{bond}}} remains a proxy. We will revise the main text to make this point clear.

[W1-2&Q1&W3&Q3-2] Report the exact number of samples that pass the stereochemical validity filter starting from all raw generated outputs, and the number of targets that obtain at least one chemically acceptable cyclic peptide candidate, and the per-target unconditional joint yield.

We counted the number of surviving samples by applying the following filters sequentially. CPSea_PDB at the top is the test set. DiffPepBuilder is omitted from the table because it generates only backbone and CβC_{\beta} atoms and therefore does not provide the complete side chains required for these filters.

Modelnumber of samplesF1F2F3F4F5F6F7F8F9Result 1Result 2
CPSea_PDB8484848484847070707070-
PepFlow840059058714400000000.000
PepGLAD (raw)840016261619800000000.000
PepGLAD (ideal.)84001371656100000000.000
Proteina-Complexa840016821266000000000.000
APCyc (raw)8400752718000000000.000
APCyc (c*)840012561202000000000.000
CP-Composer84003314000000000.000
LaCPlow-Base840021662150310206534343434126‾\underline{26}0.488‾\underline{0.488}
LaCPlow-Cyc840051235096764460161122122122121571.440

When generating 100 samples per target, only LaCPlow obtains samples that pass all the criteria. Updated: with the additional baselines reported in our consolidated comment, Proteina-Complexa (fine tuning) and PocketXMol also obtain passing samples. Please see that comment for the complete table.

F1 : + PdbondP_{d_{\mathrm{bond}}}

F2 : + backbone continuity

F3 : + bond lengths within bounds

F4 : + bond angles within bounds

F5 : + no internal clash

F6 : + no twisted peptide bonds

F7 : + energy ratio

F8 : + all-L chirality

F9 : + dG< 0

Result 1 : number of targets with at least one valid sample

Result 2 : per-target mean % (unconditional joint yield)

[Filter notes]

backbone continuity : ∣C(i)−N(i+1)∣<=2.0A˚|C(i)-N(i+1)| <= 2.0 \text{Å}

bond lengths within bounds, bond angles within bounds, no internal clash: calculated using default PoseBusters settings.

no twisted peptide bonds: candidates with a non-planar peptide bond are excluded using MolProbity; both cis and trans conformations are permitted.

energy ratio : PoseBusters rebuilds 50 conformers of the same cyclic peptide from its InChI, takes their mean UFF energy and compares the pose against it. UFF(pose) / mean UFF(50 ETKDG conformers rebuilt from the InChI) <= 100

[W2&Q2] Compare the diversity of valid samples between LaCPlow and the baselines under the same sampling budget.

PepGLAD, DiffPepBuilder, PepFlow, APCyc, and Proteina-Complexa were retrained on the same CPCore split. CP-Composer uses the authors’ released checkpoint trained on linear peptides. PepGLAD (ideal.) includes side-chain idealization, while APCyc (c*) receives the ground-truth cyclization class. Proteina-Complexa’s inference-time optimization is excluded.

We set the sampling budget to the time that the fastest model, APCyc, needs to generate 100 samples per target (88 seconds). Under this budget, we filtered the samples generated by each model using the criteria listed below and compared LaCPlow with the baseline models. The number of generated samples per target is rounded to the nearest integer.

Modelgenerated samples/targetsamples retained after filtering/targetstructural diversity (clusters/target)sequence diversity (clusters/target)
PepFlow281.281.271.28
DiffPepBuilder120.10.10.1
PepGLAD (raw)67000
PepGLAD (ideal.)670.040.040.04
Proteina-Complexa90.150.140.15
LaCPlow-Base449.66‾\underline{9.66}4.43‾\underline{4.43}9.61‾\underline{9.61}
APCyc (raw)100000
APCyc (c*)100000
CP-Composer57000
LaCPlow-Cyc4422.54722.09

Under this same budget and criteria, LaCPlow generated a larger number of distinct structural and sequence clusters compared to baselines.

[Filter]

dG<0 ∧ dbondd_{\mathrm{bond}} window ∧ CβC_\beta window ∧ Rama allowed ≥ 0.80 ∧ Rama favoured ≥ 0.70

[Q3-1] Report target-level means and confidence intervals, treating the 84 targets as independent units rather than all 8,400 generated candidates.

As specified in Section 4.1, reported values in the paper are first aggregated within each target and then averaged uniformly over the 84 targets. Here, we also report point estimates with 95% bootstrap confidence intervals. Point estimates follow the same target-wise aggregation.

ModelPdbondP_{d_\mathrm{bond}}↑PCβP_{C_\beta}↑Chirality ↑Rama allowed ↑Rama favored ↑
PepFlow7.0 [6.4, 7.7]93.9 [92.8, 94.8]100.0 [100.0, 100.0]92.4 [92.2, 92.6]81.7 [81.5, 82.0]
DiffPepBuilder1.4 [1.1, 1.8]95.0 [93.9, 96.0]100.0 [100.0, 100.0]91.0 [90.2, 91.7]78.2 [76.9, 79.4]
PepGLAD (raw)19.4 [18.4, 20.3]96.1 [95.1, 97.0]49.9 [49.5, 50.3]47.3 [46.9, 47.7]29.0 [28.7, 29.4]
PepGLAD (ideal.)16.3 [15.6, 17.1]81.9 [80.4, 83.4]100.0 [100.0, 100.0]43.8 [43.4, 44.2]25.7 [25.4, 26.1]
Proteina-Complexa20.0 [17.5, 22.6]96.3 [94.6, 97.7]99.9 [99.9, 99.9]73.3 [72.2, 74.4]51.9 [50.6, 53.2]
LaCPlow-Base25.8 [24.8, 26.7]95.5 [94.5, 96.3]100.0 [100.0, 100.0]95.3 [95.1, 95.5]83.5 [83.0, 83.9]
MuCO2.8 [2.4, 3.1]78.7 [77.5, 79.9]50.3 [50.0, 50.6]96.2 [95.6, 96.8]89.4 [88.5, 90.2]
APCyc (raw)9.0 [8.3, 9.6]74.4 [73.0, 75.8]50.1 [49.7, 50.4]43.3 [43.0, 43.7]24.9 [24.6, 25.2]
APCyc (c*)15.0 [14.3, 15.7]85.2 [84.2, 86.1]49.9 [49.6, 50.2]43.2 [42.8, 43.6]25.0 [24.6, 25.3]
CP-Composer0.4 [0.2, 0.6]74.7 [73.2, 76.2]50.2 [49.9, 50.6]39.1 [38.5, 39.8]18.6 [18.0, 19.1]
LaCPlow-Cyc61.0 [59.1, 62.7]93.9 [92.9, 94.8]100.0 [100.0, 100.0]94.5 [94.3, 94.7]83.5 [83.1, 83.9]
ModelEnergy ↑dG ↓Vina ↓GRAVY ↑
PepFlow87.3 [86.3, 88.2]−15.9 [−16.5, −15.2]−3.98 [−4.10, −3.86]59.3 [57.7, 60.7]
DiffPepBuilder95.5 [94.3, 96.5]−21.4 [−22.0, −20.8]−4.01 [−4.12, −3.90]90.5 [86.6, 93.9]
PepGLAD (raw)91.5 [90.5, 92.3]−18.1 [−18.7, −17.5]−3.55 [−3.64, −3.46]85.9 [84.9, 86.9]
PepGLAD (ideal.)90.9 [89.9, 91.9]−17.8 [−18.4, −17.2]−3.56 [−3.65, −3.47]86.1 [85.1, 87.2]
Proteina-Complexa85.3 [82.2, 87.9]−18.4 [−19.3, −17.4]−3.70 [−3.83, −3.58]91.5 [89.6, 93.2]
LaCPlow-Base97.2 [96.5, 97.8]−22.0 [−22.9, −21.1]−4.11 [−4.22, −4.01]91.1 [90.1, 92.1]
APCyc (raw)92.2 [91.4, 93.0]−16.2 [−16.7, −15.7]−3.08 [−3.15, −3.01]78.5 [77.1, 80.0]
APCyc (c*)92.0 [91.2, 92.8]−16.1 [−16.6, −15.6]−3.12 [−3.19, −3.05]86.9 [85.8, 87.9]
CP-Composer71.3 [68.6, 73.9]−13.4 [−13.9, −12.9]−2.49 [−2.58, −2.40]69.2 [68.0, 70.5]
LaCPlow-Cyc97.0 [96.2, 97.6]−20.6 [−21.4, −19.7]−4.19 [−4.29, −4.10]86.8 [85.6, 88.0]
Modeldiversity ↑novelty ↑scRMSD H2T ↓scRMSD C2C ↓
PepFlow0.50 [0.46, 0.54]0.572 [0.566, 0.579]3.52 [3.43, 3.62]3.50 [3.41, 3.60]
DiffPepBuilder0.17 [0.13, 0.20]0.515 [0.506, 0.525]3.12 [3.01, 3.22]-
PepGLAD (raw)0.62 [0.56, 0.67]0.612 [0.606, 0.618]3.69 [3.65, 3.75]3.66 [3.56, 3.77]
PepGLAD (ideal.)0.62 [0.57, 0.67]0.608 [0.602, 0.614]3.73 [3.67, 3.78]3.65 [3.54, 3.76]
Proteina-Complexa0.12 [0.10, 0.16]0.508 [0.501, 0.519]2.52 [2.30, 2.73]2.68 [2.46, 2.88]
LaCPlow-Base0.20 [0.16, 0.24]0.541 [0.533, 0.549]3.52 [3.41, 3.59]3.63 [3.53, 3.73]
APCyc (raw)0.56 [0.51, 0.61]0.625 [0.619, 0.630]3.85 [3.79, 3.91]3.94 [3.78, 4.10]
APCyc (c*)0.56 [0.51, 0.61]0.625 [0.619, 0.630]3.88 [3.81, 3.95]3.87 [3.75, 4.00]
CP-Composer0.94 [0.93, 0.96]0.622 [0.620, 0.632]3.83 [3.79, 3.86]-
LaCPlow-Cyc0.22 [0.17, 0.26]0.534 [0.527, 0.542]3.48 [3.38, 3.56]3.60 [3.52, 3.69]

Comments on the paper

2 comments
AuthorsOpenReview ↗

Consolidated Evaluation

Dear reviewers vk4B, ZcxP, and HQq7, we sincerely thank you for the constructive feedback. Since several of the requested architectural comparisons and statistical analyses can be addressed together and offer deeper insight, we address them in a single consolidated evaluation.

[vk4B W1-2&Q1&Q3-2 / ZcxP Q1 / HQq7 Q1] Staged filtering from all raw generations, unconditional joint yield, and target coverage, now including PocketXMol (ZcxP) and the fine-tuned Proteina-Complexa (HQq7).

We used the authors’ released checkpoints for CP-Composer, PocketXMol and Proteina-Complexa(fine tuning), for Proteina-Complexa(fine tuning), we fine-tuned the model using the LoRA fine-tuning protocol from the official repository. CPSea_PDB at the top is the test set.

Modelnumber of attemptsGeneration successF1F2F3F4F5F6F7F8F9R1R2R3R4R5R6
CPSea_PDB848484848484847070707070-----
PepFlow8400840059058714400000000.000---0.000
PepGLAD (raw)8400840016261619800000000.000---0.000
PepGLAD (ideal.)840084001371656100000000.000---0.000
Proteina-Complexa (from scratch)8400840016821266000000000.000---0.000
Proteina-Complexa (fine tuning)840084001384138067719714914814814814644‾\underline{44}1.73843.862.31.432‾\underline{1.432}0.750‾\underline{0.750}
APCyc (raw)84008400752718000000000.000---0.000
APCyc (c*)8400840012561202000000000.000---0.000
CP-Composer840084003314000000000.000---0.000
PocketXMol8400752770624911489548724582386338483737290.44035.148.6‾\underline{48.6}1.2760.440
LaCPlow-Base84008400216621503102065343434341260.48831.7‾\underline{31.7}56.11.1920.369
LaCPlow-Cyc8400840051235096764460161122122122121571.440‾\underline{1.440}20.737.21.5961.083

LaCPlow-Cyc achieves both the broadest target coverage and the highest mean number of distinct passing structural clusters. PocketXMol’s results further show that a high closure-distance pass rate does not necessarily translate into a higher joint yield when the desired stereochemistry is also required. Although Proteina-Complexa(fine tuning) produces more passing samples overall, LaCPlow-Cyc yields more structural clusters per target, with less concentration in a small subset of targets. Its advantage therefore extends to obtaining structurally distinct candidates that jointly satisfy the quality criteria across a broader range of targets. Furthermore, among the models trained from scratch on the cyclic peptide dataset CPCore alone, LaCPlow is the only one with samples that pass the full filtering pipeline. We will incorporate this analysis into the revised manuscript.

F1 : + PdbondP_{d_{\mathrm{bond}}}

F2 : + backbone continuity

F3 : + bond lengths within bounds

F4 : + bond angles within bounds

F5 : + no internal clash

F6 : + no twisted peptide bonds

F7 : + energy ratio

F8 : + all-L chirality

F9 : + dG< 0

Results are:

R1 : number of targets with at least one valid sample

R2 : per-target mean % (unconditional joint yield). 100 attempts per target in the denominator.

R3 and R4 : Percentage of final passing candidates contributed by the top 5 / top 10 targets, respectively. Targets are ranked separately for each model by their final passing counts.

R5 : mean number of distinct structural clusters per successful target, averaged only over the R1 targets that yield at least one fully passing candidate.

R6 : the same structural cluster count averaged over all 84 targets, counting a target with no passing candidate as zero.

[Filter notes]

backbone continuity : ∣C(i)−N(i+1)∣<=2.0A˚|C(i)-N(i+1)| <= 2.0 \text{Å}

bond lengths within bounds, bond angles within bounds, no internal clash: calculated using default PoseBusters settings.

no twisted peptide bonds: candidates with a non-planar peptide bond are excluded using MolProbity; both cis and trans conformations are permitted.

energy ratio : PoseBusters rebuilds 50 conformers of the same cyclic peptide from its InChI, takes their mean UFF energy and compares the pose against it. UFF(pose) / mean UFF(50 ETKDG conformers rebuilt from the InChI) <= 100

Reviewers vk4B(W2&Q2) and ZcxP(W2&Q2) requested a comparison of the diversity of filtered samples under the same sampling budget. We therefore report the retained candidates and the distinct structural and sequence clusters per target under a budget of 88 seconds. Candidates must satisfy the filter below, which is less strict than the cumulative filter of the previous table.

[Filter]

dG<0 ∧ dbondd_{\mathrm{bond}} window ∧ CβC_\beta window ∧ Rama allowed ≥ 0.80 ∧ Rama favoured ≥ 0.70

Modelgenerated samples/targetsamples retained after filtering/targetstructural diversity (clusters/target)sequence diversity (clusters/target)
PepFlow281.281.271.28
DiffPepBuilder120.10.10.1
PepGLAD (raw)67000
PepGLAD (ideal.)670.040.040.04
Proteina-Complexa(from scratch)90.150.140.15
Proteina-Complexa(fine tuning)91.450.971.12
APCyc (raw)100000
APCyc (c*)100000
CP-Composer57000
PocketXMol130.10.10.1
LaCPlow-Base449.66‾\underline{9.66}4.43‾\underline{4.43}9.61‾\underline{9.61}
LaCPlow-Cyc4422.54722.09

LaCPlow’s lower overall structural diversity score in Table 1 does not translate into fewer distinct candidates meeting these quality criteria. Under this budget, LaCPlow yields more structural and sequence clusters than every baseline, including Proteina-Complexa (fine tuning) and PocketXMol. PocketXMol's generation failures (873 of 8400) are not charged against its time budget. This table will also be included in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.