acceptodds
Under review as a conference paper at ICLR 2027

SupraTITO: Transferable Generative Molecular Dynamics for Supramolecular Systems

Abstract

Peptide sequence governs the structures and dynamics that emerge during supramolecular assembly, but predicting them requires resolving slow collective processes among many interacting molecules. Molecular dynamics (MD) provides microscopic insight, yet long assembly times and the vast peptide sequence space make systematic exploration costly. We introduce SupraTITO, a transferable generative molecular dynamics (GenMD) framework for supramolecular dynamics, demonstrated for peptide self-assembly. SupraTITO learns transferable implicit transfer operators (TITO) conditioned on peptide sequence, molecular topology, and periodic geometry, allowing configurations to be propagated over intervals much longer than an MD integration step. On a comprehensive dipeptide benchmark, SupraTITO generalizes to held-out sequences, reproduces sequence-dependent structures and dynamics, and maintains molecular integrity over long rollouts. Compared with direct ensemble prediction trained on the same trajectory data, it produces more accurate assembly structures while also recovering their temporal evolution. SupraTITO further generalizes across peptide concentrations, including dilute conditions absent from training. These results extend transferable GenMD to collective dynamics in periodic supramolecular systems and related processes beyond peptide assembly.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Molecular Simulation, Generative Models, Coarse-Graining, Peptide Assembly, Supramolecular Systems

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_emsbOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents SupraTITO, an implicit transfer operator for generating trajectories of supramolecular systems, especially peptide self-assembly. Conditioned on the peptide sequence, molecular topology, and periodic geometry, the model predicts the minimum-image site displacements of molecules over a large time lag, from which the current state can be recursively updated. On the dipeptide benchmark consisting of 400 systems, SupraTITO generalizes well to held-out systems and even diverse peptide concentrations unseen during training.

Critical strengths
  • Compared with prior work that primarily focuses on single-molecule MD, peptide-assembly dynamics represent a relatively novel and important direction. SupraTITO provides a simple yet effective modeling approach to this problem, and the experimental results are also quite promising.
  • The overall experimental setup is reasonable and comprehensive, including evaluations of generalization across peptide sequences, a broad set of metrics, and appropriate baseline comparisons. I am particularly interested in the model's generalization to peptide concentrations that are unseen during training, as this provides strong evidence that the model has learned intermolecular interactions rather than merely memorizing the training data.
Critical weaknesses

None

Questions for authors
  1. I think the manuscript could be reorganized to improve readability. Currently, the Conclusion section is crowded with several figures and tables, which makes the layout somewhat cluttered. In particular, some less essential details in the Methods section, such as Eqs. (15–17), could be moved to the Appendix.
  2. Lines 190–193 state that the method adopts a two-site-per-molecule representation. It is feasible because the dipeptide system can be coarse-grained into two CA atoms. If the method is extended to peptide sequences containing three or more residues, however, the current representation may no longer be directly applicable. This therefore seems to limit the generalizability of the representation. Please correct me if my understanding is inaccurate.
  3. The experiments appear to focus exclusively on peptide self-assembly systems. I would like to know whether the current model can also be applied to cross-assembly systems, for example, mixtures of two or more different dipeptides at varying concentrations.
  4. The current experiments appear to model systems with implicit solvent. I would also like to know whether the model can support systems with explicit water molecules. I understand that running additional experiments may require more time than is available during the rebuttal period, so a feasibility analysis alone would also be helpful.
4 comments
AuthorsOpenReview ↗

Response to Reviewer emsb, Part 1/3

Thank you for the positive assessment and helpful suggestions. We have added coassembly tests with the submitted model and a separate mixture-training experiment, and clarify below how the formulation relates to longer peptides and explicit solvent. We address your questions in order below.

Clearer presentation and figure placement

Question 1

We agree that the organization and layout can be improved. We are incorporating all reviewers' feedback and will upload a revised manuscript soon. With the additional page available for revision, we will improve figure placement around the Conclusion and move Eqs. (15-17) to the Appendix.

Extension beyond two sites per molecule

Question 2

We understand the confusion: the paper expresses the molecule-consistent displacement using a center shift and a change in a single bond vector, which is specific to dipeptides. The underlying idea is more general: choose the periodic image for the molecule as a whole, rather than independently for each site.

For example, a longer peptide crossing a box boundary can appear split between opposite sides of the box. Using its bonded connectivity, we first reconstruct the connected molecule at both ends of the transition. We then assign all sites the same displacement of the molecular center and add each site's change in position relative to that center. This avoids inconsistent box-length shifts between bonded sites while allowing the peptide to rotate and change conformation; it does not treat the molecule as rigid.

Let mm index molecules and nmn_m denote the number of sites in molecule mm. Let yt,i(m)\mathbf{y}_{t,i}^{(m)} be the reconstructed coordinates of site ii at time tt and gt(m)=nm−1∑i=1nmyt,i(m)\mathbf{g}_t^{(m)}=n_m^{-1}\sum_{i=1}^{n_m}\mathbf{y}_{t,i}^{(m)} the molecular geometric center. The displacement target for each site i=1,…,nmi=1,\ldots,n_m is

dt,τ,i(m)=MIC⁡L(gt+τ(m)−gt(m))+(yt+τ,i(m)−gt+τ(m))−(yt,i(m)−gt(m)).\mathbf{d}_{t,\tau,i}^{(m)} = \operatorname{MIC}_L(\mathbf{g}_{t+\tau}^{(m)}-\mathbf{g}_t^{(m)}) + (\mathbf{y}_{t+\tau,i}^{(m)}-\mathbf{g}_{t+\tau}^{(m)}) - (\mathbf{y}_{t,i}^{(m)}-\mathbf{g}_t^{(m)}).

For a dipeptide, the two positions relative to the center are minus and plus half the bond vector, recovering the formula in the paper. For a longer peptide, we can use each site's position relative to the molecular center instead. This construction is not restricted to one bead per residue: sites can instead represent all backbone and side-chain beads defined by a coarse-grained force field such as Martini, or individual atoms when an all-atom force field is used. The shared networks can process the additional sites and their connections. We will clarify the construction in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer emsb, Part 2/3

Coassembly of different dipeptides

Question 3

We agree that coassembly is a very interesting test and thank the reviewer for bringing this up. Our architecture already supports inference for mixtures: each site carries its own amino acid and molecular role features, while pair features describe connectivity and geometry across all peptides. Different peptide species can therefore be included in the same configuration. The original experiments focused on self-assembly because training used only homogeneous systems, but we were keen to test this extension directly.

We prepared 11 external coassembly tests at the training concentration and peptide number, isolating transfer to new compositions. The test cases cover different mixing ratios, binary and ternary mixtures of peptides with contrasting assembly tendencies, differences in peptide chemistry, and reversed residue order. The test scenarios and compositions list these cases and the training membership of their component sequences.

The coassembly table and mean and variability comparison summarize late SASA across these mixtures. Using the frozen submitted model trained only on self-assembly, the blue curves in the SASA comparison for all 11 mixtures show close agreement for several mixture families, including FF/YY, FF/FI, and FG/GF. The main differences occur in mixtures containing GG: the final snapshots show more GG incorporated into the aggregate than in MD, consistent with the lower predicted SASA. A plausible explanation is that homogeneous training does not expose the model to the mixed molecular environments that determine whether GG joins an aggregate or remains dispersed.

We therefore trained a separate model on the original 320 self-assembly trajectories together with 20 mixtures of ten peptide species each, without changing the architecture. The mixture species were selected from the original training partition to cover diverse peptide pairings and chemical environments; the full training mixture list gives their compositions. None of the 11 test compositions are included in training. The orange curves in the same SASA comparison show substantially improved agreement for the GG mixtures, bringing their solvent exposure close to MD. In the snapshot comparison, more GG remains outside the aggregate, closer to the balance between assembled and dispersed peptides in MD. These results extend the demonstrated scope of SupraTITO beyond homogeneous self-assembly to multicomponent peptide systems.

Explicit solvent in the reference MD and learned model

Question 4

Thank you for raising this point. Our reference MD includes explicit Martini water and counterions, but SupraTITO learns only the peptide motion from these simulations, without generating solvent coordinates.

We suppose you are also asking whether SupraTITO could generate explicit water. We think this is possible with suitable training data and solvent features, but expensive with the current global attention, whose cost scales quadratically with the number of sites. A more targeted extension could generate water near peptide interfaces or within cavities, where hydration may influence assembly morphology, while treating bulk solvent separately. A practical alternative is to generate peptide configurations, reconstruct the omitted peptide beads, then solvate and run short MD simulations for relaxation.

A local architecture is a promising direction for joint peptide and water generation. For example, AquaGen (Bengio et al., 2026) uses graph message passing over nearby atoms to generate explicit solvent configurations under periodic boundaries, demonstrating this approach for equilibrium sampling. Extending local models to assembly dynamics is therefore worth exploring, although their accuracy and practical scalability for this task still need to be tested.

AuthorsOpenReview ↗

Response to Reviewer emsb, Part 3/3

Additional results: Architecture ablations and baselines

We also added architecture ablations in response to Reviewer 7Yyr. The ablation table supports the molecule-consistent displacement construction discussed above, together with explicit bond features and pair geometry evaluated from physical positions. These controls use the same training split and budget and provide further evidence for the design choices in SupraTITO. Details are provided in our response to Reviewer 7Yyr.

Following Reviewer A5mu's suggestion, we also added baselines including nearest sequence retrieval and observable prediction from residue descriptors and protein language model embeddings. The baseline comparison shows lower mean prediction error for SupraTITO. Further details are provided in our response to Reviewer A5mu.

I appreciate the authors' thorough experiments and detailed explanations, which have satisfactorily addressed my concerns. I will maintain my positive score and believe this paper deserves acceptance at the conference.

Reviewer_A5muOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3(+1)
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper proposes SupraTITO, a generative MD framework that extends transferable ITO to supramolecular systems, demonstrated on peptide self assembly. SupraTITO conditions on peptide sequence, molecular topology, and periodic geometry, SupraTITO learns the distribution of finite-lag displacements of coarse-grained residue site. The method is trained on 320 dipeptides and evaluated on 80 held-out sequences by structural and dynamical observables.

Critical strengths

1. Extension of transferable ITO to supramolecular systems

To the best of my knowledge, this seems to be the first paper extending generative MD and ITO to a supramolecular system, where periodic boundaries have to be considered. The consistent displacement is simple and exact, keeping molecules intact.

Critical weaknesses

1. Lack of baselines to compare with

The only baseline is DEP constructed by the authors, while I understand there is no prior method attempting to assemble trajectories, and thus comparisons to existing models are not expected. However, simple sequence-to-observable baselines remain feasible and would clarify what the dynamical formulation adds. For example, regressing last-time observables with ESM embeddings, similar to PLaTITO's use of protein language model representations, or per-residue descriptors. Such baselines may not be able to produce trajectories, but would strengthen the paper's claim on how well SupraTITO performs and how hard the task is.

2. Distribution of observables and interpretation of correlation

It seems that the observables appear strongly clustered, many sequences are fully assembled or fully dispersed, with relatively few intermediate cases. Under such a distribution, the Pearson correlation can be high just from separating the dispersed and assembled groups, while ordering within each group matters little. Could the authors (i) show the distribution of each observable over the training and held-out sets, and (ii) report agreement within groups, e.g., a subset of intermediate sequences where the authors note the largest errors, would address this.

[1] PLaTITO, https://arxiv.org/abs/2602.11216

Questions for authors

1. Construction of the held-out split

The held-out sequences are selected to balance amino-acid coverage and the distribution of the late-time MD SASA. First, since these quantities are outcomes of the trajectories being predicted, was a split based only on sequence-level properties considered? Second, because balancing keeps held-out sequences within the training range of assembly behavior, how does transfer perform when, for instance, the most strongly assembling sequences are chosen as the held-out? Finally, how was the balancing performed in detail?

3 comments
AuthorsOpenReview ↗

Response to Reviewer A5mu, Part 1/2

Thank you for recognizing our contribution and for the suggestions on evaluation. We have added baselines for observable prediction, comparisons within assembly groups, and two additional training runs testing a split based only on sequences and extrapolation to systems with the lowest SASA.

Baselines

Critical Weakness 1

We agree that sequence-to-observable models help establish what trajectory generation adds. We therefore evaluated nearest-sequence retrieval and ridge, gradient-boosting and separate/joint MLP regressors. Their inputs span residue identity, physicochemical (Atchley) and MARTINI interaction descriptors, as well as frozen protein-language-model embeddings, not only from ESM-2 but also from PepBERT. These controls test whether related training sequences or direct sequence-based prediction already explain the agreement we report.

All comparisons use the original 320/80 sequence split, and the same six late-window MD targets. We carefully chose regressor complexity and hyperparameters via training-only five-fold cross-validation, keeping reverse sequences together and fitting preprocessing within each fold. We searched over multiple model families and extended capacity grids rather than giving these a transformer-sized parameter budget, given the small amount of data present without trajectory information (320 sequence-level targets). In the extended MLP grid, four 128-unit layers were selected over three 256-unit layers in all 15 feature/seed searches. These results suggest diminishing returns for the already strong representations, consistent with the small number of labeled sequences.

The CV-selected sequence predictor achieves 7.52% mean normalized MAE, compared with 6.56% for DEP and 2.50% for SupraTITO. SupraTITO's advantage therefore extends beyond the generative baseline to direct property prediction. Baseline comparison, methods and selection.

We additionally test whether the starting observables suffice: a regressor given the six frame-0 observables and sequence descriptors achieves 5.67%, while sequence-conditioned stochastic observable-space dynamics achieves 5.54%. The latter learns transitions from training trajectories and is selected by late-window prediction error after rollout. These controls extend the comparison beyond sequence prediction and unconditional configuration generation, showcasing that SupraTITO improves observable prediction while also generating the underlying molecular configurations. Scores and intervals.

All scores average predictions across seeds before evaluation; MAE is normalized by each observable’s held-out MD range and averaged over six observables. Intervals use 5,000 reverse-pair bootstrap resamples.

AuthorsOpenReview ↗

Response to Reviewer A5mu, Part 2/2

Observable agreement and sequence selection

Critical Weakness 2 and Question 1

We agree that these analyses are useful for interpreting the overall correlations. As requested, we now show the distribution of each observable over the training and held out sets in the observable distribution plot. Both sets reflect the broad pattern in the full dataset, where strongly assembled systems are more common than intermediate systems; this imbalance is not specific to the held out partition. We also report agreement within groups, including the subset of intermediate sequences, in the expanded within-group comparison. With the comparison models fixed before subgroup evaluation, SupraTITO has the lowest mean normalized error in each assembly group and the highest mean correlation among the 19 intermediate sequences. This comparison includes sequence prediction, frame-0 prediction and observable-space dynamics alongside DEP, showing agreement within assembly regimes rather than relying only on separating dispersed and assembled systems.

The original split was designed to test new peptide combinations across the observed assembly range. It uses MD outcomes to balance the evaluation set: reverse sequences remain together, each amino acid has four partners held out, and homodipeptides remain in training. Selection balances late SASA and relative SASA reduction averaged over reverse pairs, using five bins for each distribution alongside groups defined by chemical properties. It also penalizes SASA values outside the corresponding amino acid's training range. We will describe this procedure in detail in the Appendix.

To test whether the results depend on outcome balancing, we retrained the same architecture with the same budget on a random split (seed 12345) based only on peptide sequences. We randomly select pairs of reverse sequences for evaluation without using MD observables, keeping XY and YX in the same partition and retaining 320 training and 80 evaluation sequences. The random split comparison plot shows close agreement across the assembly range, and the table of all six observables shows errors comparable to the original split. Although the test sequences differ, this control demonstrates that good prediction does not require balancing based on MD outcomes.

Following your suggestion to hold out the strongest assemblers, we also retrained with the sequence groups having the lowest late SASA held out, keeping reverse pairs together and all other settings unchanged. This places 76 of the 80 evaluation sequences below the remaining training SASA range. The lowest SASA holdout plot and corresponding score table show that SupraTITO predicts compact assemblies beyond that range. This directly tests extrapolation toward low SASA assemblies, rather than requiring the assembly behavior of evaluation sequences to remain within the training range.

The split membership table lists the evaluation sequences for the original split and both new controls.

Additional results: Architecture ablations and coassembly

We also added architecture ablations, discussed in our response to Reviewer 7Yyr. The ablation table supports the use of molecule-consistent displacements, explicit bond features, and pair geometry evaluated from physical positions. Following Reviewer emsb's suggestion, we further tested coassembly: the coassembly comparison shows transfer to several mixtures with the submitted model and substantial improvement for GG mixtures after adding 20 mixture trajectories to training. None of the 11 test compositions are included in training. These results broaden the evidence for transfer beyond homogeneous peptide systems; details are provided in our response to Reviewer emsb.

I thank the authors for all the additional experiments on simple sequence-based baselines, observable distribution, and sequence-split results, along with a detailed explanation; all of my concerns have been resolved! I have raised my score accordingly.

Reviewer_7YyrOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper extends transferable learned "time-jump" MD models (implicit transfer operators) from single molecules to many-molecule systems: 100 coarse-grained dipeptides self-assembling in a periodic box. A flow-matching model, trained on time lags from 2.5 ps to 2.5 ns, samples how all the molecules move over a given lag. Each molecule's motion is split into a minimum-image centre shift plus a bond-vector change, and pair geometry is recomputed at every flow step to handle periodic boundaries. At inference the model is run only at 2.5 ns, repeatedly, to build 1.25 µs trajectories. It is trained on 320 of the 400 dipeptide sequences and tested on the other 80, as well as at other concentrations and system sizes.

Critical strengths

This is a credible first step from single molecules to many-molecule assembly, and the structural results look solid. The held-out evaluation is careful: reverse sequences are held out together, and the two metric constants (SASA bead radius, contact cutoff) were fitted on training data only. Results are shown per sequence. Time-averaged structure agrees well, least so for intermediate-assembly sequences (Fig. 2c), and single-frame morphology agrees less well (Fig. 7). The model beats a baseline trained on the same trajectories (but only their final 125 ns, with no current-state input), with about 1.7–2.6× lower error on all six structural observables (Fig. 5b). That baseline's capacity and tuning aren't shown to match.

Critical weaknesses
  1. The dynamics claims need more evidence. Training covers lags from 2.5 ps to 2.5 ns, but all results use only 2.5 ns, and there is no check that shorter steps compose consistently, although the paper itself says this should be tested (Sec. 2.3). No assembly time or rate is extracted and compared with MD. The main timescale comparison is a contact-memory measure (Fig. 3d). It is coarse (2.5 ns output spacing), the authors call it a diagnostic rather than a rate, and it modestly underestimates some of the longest-lived contacts. In Fig. 3c the model also assembles faster than MD. The reference is a single MD run per sequence, as the authors acknowledge, so model error can't be separated from run-to-run variation.
  2. The scope of transfer is narrower than the framing, and controls are missing. Every amino acid appears in training, so held-out sequences are new combinations, and the authors' own coverage study shows performance drops quickly with fewer training sequences (Fig. 4). In the concentration study (Fig. 6), five of the six sequences are in the training set and four are nearly flat, so only QI tests a new sequence at a new concentration. Size transfer is inconsistent (Fig. 25: FF at 50 peptides and QI at 200 peptides are clearly off). The authors report this openly (Sec. 5.3, App. H.2), but the abstract's broader generality claim should be tempered to match. There is no simple baseline, such as using the nearest training sequence, and no ablation of the periodic-handling components.

Overall: If lag-consistency or replicate-MD tests show the dynamics are unreliable, I would lower my rating.

Questions for authors
  1. Can you show that shorter lags compose consistently with the 2.5 ns step, for example rollouts at 250 ps or 1 ns compared on the same statistics?
  2. How do assembly times compare with MD across the held-out sequences?
  3. For a few sequences, how large is the variation between independent MD runs?
  4. How does a nearest-training-sequence baseline do on the late-time observables?
  5. Can you add held-out sequences in the intermediate-assembly regime to the concentration study? What goes wrong for QI at 200 peptides?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 7Yyr, Part 1/2

Thank you for recognizing our contribution and for the suggestions to strengthen the evaluation. We have added shorter-lag rollouts, assembly-time comparisons, independent MD replicas, further concentration tests, additional prediction baselines, and periodic-handling ablations to address your questions.

Agreement across rollout lags and assembly times

Critical Weakness 1, Questions 1-2

We agree that comparisons across rollout lags and assembly times are useful tests. We therefore added 250 ps and 1 ns rollouts using the same checkpoint and setup as the reported 2.5 ns results, with five rollouts per sequence. As shown in the lag comparison, all three rollout lags give broadly similar agreement with MD and substantially outperform DEP across all six observables, supporting consistency across rollout lags.

We use 2.5 ns because it is the largest training lag and requires fewer rollout steps for a given trajectory duration. In practice, using the largest trained lag is a natural choice for efficient rollout when prediction accuracy is maintained. At smaller lags, errors can accumulate over more autoregressive rollout steps. Small lags can also introduce errors when memory from unresolved degrees of freedom violates the Markovian assumption (Sec. 2.3). Both effects may contribute to the slightly higher errors at 250 ps. Training at larger physical lags could further reduce rollout cost.

For dynamics, beyond contact memory, the paper also reports the time evolution of SASA (Fig. 8) and largest-cluster fraction (Figs. 3a-c and 9), together with cluster-fraction autocorrelations (Fig. 11). Nevertheless, we have added assembly-time comparisons to address your question directly. Both the 50% threshold comparison and 75% threshold comparison (the first time a cluster contains at least 50% or 75% of the peptides, respectively, sustained for 10 ns) show broadly comparable assembly timescales between MD and SupraTITO, now including the additional MD replicas where available.

Variation across independent MD replicas

Critical Weakness 1, Question 3

To assess how much assembly varies between MD runs, we generated two additional independent replicas for each of 15 sequences spanning weak, intermediate, and strong aggregation. The 50% assembly-time comparison and 75% assembly-time comparison show substantial variation in intermediate systems, including assembly thresholds reached in only some MD replicas.

Regarding the faster assembly you noted for YP in Fig. 3c, the representative trajectories (panel c) now show that SupraTITO's cluster-growth curve overlaps the variation across independent MD replicas. The 50% and 75% assembly-time comparisons further show comparable assembly timescales for YP, with earlier or later assembly depending on the sequence. Together, these results suggest that MD variability contributes to the original difference, rather than the model uniformly accelerating assembly.

AuthorsOpenReview ↗

Response to Reviewer 7Yyr, Part 2/2

Nearest-training-sequence baseline

Critical Weakness 2, Question 4

To test whether nearby training sequences already explain the predictions, we added nearest-sequence baselines using ordered Atchley and Martini descriptors, selecting neighbors without held-out outcomes. The score table shows SASA MAEs of 33.61 and 27.06 nm2nm^2, versus 5.03 nm2nm^2 for SupraTITO. Across the six observables, mean range-normalized errors are 15.25%, 12.20%, and 2.50%, respectively; the comparison plot shows these normalized errors and their uncertainty intervals. Beyond retrieval, a sequence-to-observable predictor selected by training-only grouped CV achieves 7.52%, versus 2.50% for SupraTITO. The expanded comparison also includes prediction from initial observables and learned observable-space dynamics.

To address your comment on DEP's capacity and tuning, DEP removes the current-state condition network, which is not needed for direct ensemble prediction, and uses a deeper velocity network to approximately match SupraTITO's parameter count. Both models use the same optimizer and learning rate and are trained for the same number of epochs, with the same number of sampled examples per epoch. We will add a dedicated appendix subsection detailing the DEP architecture and training setup to make this comparison clearer.

Architecture and periodic-handling ablations

Critical Weakness 2

Thank you for pointing out the missing ablations of the periodic setup. We have added a detailed comparison of our design choices using the same training split and full training budget. The ablation table shows that molecule-consistent displacements improve most assembly observables and bond-length agreement compared with applying the minimum-image convention independently to each site. Removing bond pair features substantially worsens molecular geometry and peptide-contact predictions, supporting the explicit representation of molecular connectivity.

Computing pair geometry from displacements alone instead of the evolving physical configuration worsens all six assembly observables, even though some bond and contact-memory errors are lower. This supports evaluating interactions using physical molecular positions. We also tested the periodic flow-matching extension originally proposed in our future-work discussion, implementing a molecular torus variant inspired by Riemannian flow matching (Chen and Lipman, ICLR 2024), with periodic paths for molecular centers and Euclidean paths for internal motion. It gives comparable assembly predictions but no improvement over the submitted formulation on these observables. Together, the controls support our use of molecule-consistent displacements, explicit bond features, and physical pair geometry.

Additional results: Transfer to coassembly

Following Reviewer emsb's suggestion, we also tested coassembly of different peptide species. The coassembly comparison shows that the submitted model transfers to several mixtures without retraining. Adding 20 mixture trajectories to the original training data, without changing the architecture, substantially improves the GG mixtures that were less accurately predicted. All 11 test compositions remain unseen during training. These results provide further evidence of transfer beyond self-assembly and show that a small addition of mixture data can improve predictions for new compositions, complementing the sequence coverage study. Details are provided in our response to Reviewer emsb.

Comments on the paper

1 comment

We thank all reviewers for their thoughtful feedback and constructive suggestions. The revised manuscript is now uploaded. Thank you for your time and consideration!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.