acceptodds
Under review as a conference paper at ICLR 2027

Learning Topological Representations of Protein Structure and Dynamics

Abstract

Modern protein representation models support tasks such as enzyme design and drug discovery, but their reliance on static data such as sequence and native structure limits their ability to capture the conformational dynamics that drive protein function. We investigate whether persistent homology (PH) can provide descriptors shared across diverse proteins that retain global structure, fine-grained conformational variability, and kinetically relevant information without large-scale pretraining. We introduce the masked Flood complex, i.e., an adaptation of a recently proposed simplicial complex construction, that incorporates domain knowledge to emphasize inter-residue structure at low computational cost. We then use the latter to compute PH on molecular dynamics (MD) sampled structures, vectorize the persistence diagrams into a shared coordinate system, and probe the capacity of these representations in terms of the aforementioned aspects. To assess the amount of kinetic information, we learn low-dimensional embeddings from time-lagged observations and evaluate Markov state models (MSMs) estimated from them. Using these MSMs to guide training of the recent MarS-FM generative framework improves several ensemble statistics relative to the original model. After finetuning on lower-temperature MD data and adapting the sampling procedure, the resulting model also shows promising transfer to fast folding proteins.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Topology, Persistent Homology, Representation Learning, Generative 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_gPphOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper uses Flood Complexes to represent MD trajectories, which is a sensible thing to do. It further adds a protein-based prior, which is the masking of inter-residue atoms when constructing the PH features.

Critical strengths
  • The paper uses Flood Complexes to represent MD trajectories, which is an improvement upon other PH based approaches for MD.
  • The paper also adds a protein specific masking, which makes sense, and improves the results.
Critical weaknesses

No major weaknesses. The main contribution is applying Flood Complexes to MD and adding atom masking. The masking step is very intuitive, and the use of PH is not new to MD. While it is not very novel, it works well.

Questions for authors
  1. Does the analysis show the importance of the different Gaussians (exponential structure elements) on the dynamics, and explain what each element represents?

  2. How does removing some of them affect the MSM, and can combinations of these Gaussians act as an indicator for specific dynamics?

2 comments

We thank the reviewer for the feedback and for taking the time to evaluate our work. We are pleased that the reviewer finds our approach intuitive and recognizes its effectiveness.

As a general comment regarding novelty, we want to highlight that we do not just simply apply PH in the context of studying MD. Instead, our mFlood approach introduces a method to infuse domain-specific knowledge into the simplicial complex construction itself, i.e., the very basis for all downstream applications. As we lay out in our answer to reviewer g36h's first question, this renders filtration values biologically informative, e.g., with respect to the presence of secondary protein structure and folding contacts. Overall, this enables building computationally tractable topological representations that are both structurally and dynamically informative. We furthermore, deviate from the traditional method of learning kinetic embeddings specific to each protein, and introduce a topologically-informed shared kinetic embedding space. These embeddings turn out to be kinetically more consistent than those from traditional approaches and also generalize in various ways: first, as was added in this rebuttal (see section 5.2 and D.5), they do so for the lower temperature simulations of the unseen ATLAS dataset (Vander Meersche et al., Nucleic Acid Res, 2024), identifying subtle important motions; second, the generative model training, guided by MSMs, transfers to contrasting dynamics of smaller, fast-folding proteins.

In summary, novelty can be summarized as (1) the introduction of domain knowledge at the basis of simplicial complex construction via masking, (2) computational tractability and scalability of the latter, (3) building topological representations that are both structurally and dynamically informative of proteins, (4) the learning of a shared kinetic embedding space and (5) strong empirical evidence for generalizability.

Questions

Q1:: "Does the analysis show the importance of the different Gaussians (exponential structure elements) on the dynamics, and explain what each element represents?"
Q2: "How does removing some of them affect the MSM, and can combinations of these Gaussians act as an indicator for specific dynamics?"

Assessing the importance and interpretation of individual structure elements is a good idea. We added two such analyses to the paper (see Appendix G.2).

First, to measure the importance of single structure elements on the dynamics, we shuffle the values of the corresponding feature (i.e., a single coordinate of the PH vectorization) along the time axis of the MD trajectories, while preserving all other features. We then record how much the VAMP-2 loss deteriorates. The three most important structure elements identified in this manner are located in the H2 diagrams at birth times between 3.2–3.3 Å. Looking at the points in the persistence diagrams that are closest to these structure elements, we see that they correspond to small cavities that exist in the folded domain and disappear as it unfolds (see Fig. 7d–f). Moreover, for the most important structure elements, we observe that the majority of their corresponding feature (a sum of contributions from individual voids) comes from voids that are only made visible to the filtration by masking.

While the above addresses the contribution of each structure element in terms of dynamic information, we can not rule out that the VAMPNet is capable enough to restore temporal information from a combination of the other elements. To address your second question, we assess how much of the VAMP-2 score is lost when retraining the network without particular combinations of features.

First, we observe that retraining without the H0 vectorizations, i.e., without all H0 structure elements, has almost no effect on the VAMP-2 score (0.7%), while removing H1 (14.3%) and H2 (16.8%) vectorizations results in a large drop. We conclude that important slow dynamics are primarily encoded in H1 (loops) and H2 (voids).

Second, removing the overall 32 (6.3%) or 96 (12.5%) most important structure elements identified from the shuffling experiment (across H0, H1 and H2) costs 3–5 times as much as removing the same number of random ones. Furthermore and to answer your last question, we hypothesize that combinations of elements do act as indicators of specific dynamics. More specifically, the summed value of the three most important elements is high in folded conformations and drops as the domain unfolds, closely following the fraction of native contacts along a trajectory (Fig. 7 d-f). Across validation domains, the corresponding contribution follows the fraction of native contacts with a median correlation of 0.66.

We thank the reviewer once more for their input, which has led to a stronger submission. In light of the approaching end of the discussion period, we would like to ask whether our responses have sufficiently addressed all raised questions. If so, and if the reviewer deems it warranted, we would greatly appreciate a reconsideration of our score. Otherwise, we remain happy to answer any further questions.

To briefly summarize this thread:

  1. We discussed all angles in which our approach indeed displays novelty, including that our kinetic embedding transfers to unseen datasets (see also our new results on ATLAS in the thread of reviewer g36h).
  2. It is indeed possible to attribute importance to individual structure elements. Those corresponding to voids linked to small cavities are made visible by masking and contribute most to the slow dynamics captured by the kinetic embedding (Appendix G.2).
  3. Removing the most important elements costs several times as much as removing random ones, and their combination acts as an indicator of folding (Table 27, Fig. 7).
Reviewer_MS1mOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates persistent homology (PH) as a shared representation of protein structure and conformational dynamics without large-scale representation pretraining. The authors introduce the masked Flood complex, a protein-specific adaptation of the Flood complex that uses Cα atoms as landmarks and residue-dependent masks to emphasize inter-residue interactions. The resulting persistence diagrams are vectorized into fixed-dimensional representations that can be shared across different proteins.

The proposed representations are evaluated through protein structural classification, regression of physical observables, and molecular kinetic modeling. A shared VAMPNet learns low-dimensional kinetic embeddings, from which domain-specific Markov state models (MSMs) are constructed. These MSMs are subsequently incorporated into the existing MarS-FM framework to guide generative modeling of protein conformational ensembles. Experiments on mdCATH demonstrate competitive structural representation quality, improved MSM consistency over the original Flood complex, and improvements in several generative ensemble statistics. Additional experiments investigate transfer to fast-folding proteins after adapting the model using lower-temperature molecular dynamics simulations.

Critical strengths
  1. The paper introduces a well-motivated and mathematically grounded extension of persistent homology for protein representations. The masked Flood complex incorporates residue-specific structural information while preserving a valid filtration, with additional stability analysis in the appendix. This is a substantive methodological contribution because it provides a principled approach to incorporating biological inductive biases into topology-based representations without relying on large-scale pretraining.

  2. The evaluation connects representation quality with downstream molecular dynamics applications. Beyond structural classification and observable regression, the authors assess kinetic information through MSM quality metrics and demonstrate its practical utility in the MarS-FM generative framework. The improvements in Markovianity and several ensemble statistics, supported by statistical analyses, provide meaningful evidence that the proposed representations capture information relevant to protein dynamics.

Critical weaknesses
  1. The proposed masked Flood representation does not consistently outperform alternative PH constructions. In Table 3, the original Flood complex performs similarly or better on several physical observables, while all-atom Alpha PH achieves stronger results for some regression targets. Similarly, Table 5 shows that other PH variants can outperform mFlood on individual generative metrics. These results make it difficult to determine precisely which benefits arise from the proposed residue-dependent masking. A clearer analysis of the conditions under which masking provides a meaningful advantage would strengthen the central methodological claim.

  2. The generality of the kinetic representations remains insufficiently established beyond the high-temperature unfolding regime. Most kinetic experiments are performed on mdCATH trajectories at 450 K, where unfolding dominates the slow dynamics. Although the fast-folder experiments are valuable, they additionally involve lower-temperature fine-tuning and changes to the sampling procedure. Consequently, these results do not independently establish how well the learned topological representations generalize to different dynamical regimes. More direct evidence separating representation transfer from adaptation would strengthen the generalization claims.

Questions for authors
  1. In which settings does residue-dependent masking provide a clear advantage over the original Flood complex or all-atom Alpha PH? Given the mixed results in Tables 3–5, could the authors clarify whether the main benefit is improved kinetic consistency, computational efficiency, or overall robustness across tasks?

  2. How much of the improvement on fast-folding proteins is attributable specifically to the mFlood-based kinetic representation rather than lower-temperature fine-tuning and modified sampling? Are controlled comparisons isolating these factors available from the existing experiments?

  3. The proposed representations are described as computationally efficient, but Table 1 reports approximately 91 hours of preprocessing for all mdCATH conformations, compared with 46 hours for the original Flood complex. Could the authors clarify the intended efficiency claim, particularly when the complete preprocessing cost is included?

3 comments

We appreciate the reviewer's thoughtful comments and are pleased that our methodological contribution was well received. Below, we address the concerns and questions raised in the review.

Q1

"In which settings does residue-dependent masking provide a clear advantage over the original Flood complex or all-atom Alpha PH? Given the mixed results in Tables 3–5, could the authors clarify whether the main benefit is improved kinetic consistency, computational efficiency, or overall robustness across tasks?""

While there is no PH variant that is best on every probe, mFlood yields the most robust PH representation across tasks, with not only the best overall but also the best worst-case rank (see Table 19).

That said, we consider the main benefit of masking to be kinetic consistency. The Chapman–Kolmogorov (CK) error measures whether the model's transition probabilities at one lag time predict the transitions observed at multiples of it, which holds only if the dynamics in the embedding are Markovian. π-JS measures how far away the model's estimated stationary distribution lies from the state populations of the reference MD. Relative to Flood PH, masking lowers the CK error by a third and nearly halves π-JS (see Table 4).

Our response to reviewer g36h's first question explains where this comes from. With masking, the filtration value of a simplex depends only on the residues around it, not on those forming it. This makes cavities of the folded domain visible that plain Flood would miss. The kinetic model relies most on these cavities (see Appendix G.2).

Compared with all-atom Alpha PH, mFlood PH matches its CK error while capturing more slow information (i.e., VAMP2 1.503 vs 1.463) with a slightly higher π-JS at a third of the cost. It also classifies CATH classes clearly better (79.1% vs 71.3%), whereas Alpha PH leads on regression of the radius of gyration and the RMSD.

Q2

"How much of the improvement on fast-folding proteins is attributable specifically to the mFlood-based kinetic representation rather than lower-temperature fine-tuning and modified sampling? Are controlled comparisons isolating these factors available from the existing experiments?"

The models without finetuning already provide a controlled comparison: MarS-FM (Kapuśniak et al., ICLR 2026) and MarS-FM-mFlood are trained in the same way and differ only in the features from which their MSMs are estimated. Hence, comparing them isolates the representation. MarS-FM-mFlood is significantly closer to the reference in covalent geometry and folding free energy, and in how it populates secondary structure and metastable states (see Table 7, Table 23). Both models benefit from finetuning to a similar extent, with our MarS-FM-mFlood remaining closer to the reference afterwards.

We consider finetuning necessary for the fast folders because MarS-FM is trained on dynamics of an opposite kind. In particular, mdCATH trajectories at 450 K mostly show unfolding, and MarS-FM training exploits this by mainly drawing the source conformation from the starting frame's MSM cluster. That is, it trains largely on unfolding transitions. The fast-folder data, in contrast, come from an adaptive sampling scheme that explores the folding landscape thoroughly (Majewski et al., Nat. Commun. 2023).

To further separate the effects of the adaptation steps, we started from the two (not yet finetuned) models and finetuned both for 100 epochs on the 450 K data the model was already trained with, as well as on the 348 K data. In both cases, we assess the effect of the lag, sampling and temperature change on their own and in combination. The table below lists the mean over the 12 fast folding proteins for MarS-FM / MarS-FM-mFlood (ours). As a control, continuing training without any change improves bond geometry for both models but leaves the other measures largely unchanged. Lowering the lag time, changing the sampling procedure and using lower temperature data each bring large improvements for both models, and with all three steps, 100 epochs already come close to the longer 300-epoch finetuning presented in the paper. Across all settings, our MarS-FM-mFlood remains closer to the reference than MarS-FM in covalent geometry, and in all but one setting also in folding free energy and in its population of secondary structure and metastable states. The advantage of the representation thus persists regardless of the way the models are adapted.

AdaptationBond RMSZ ↓ΔG error (kcal/mol) ↓SS JSD ↓π-JS ↓
none5.78 / 4.963.50 / 3.190.490 / 0.4610.414 / 0.392
continued training only4.99 / 4.003.41 / 3.220.497 / 0.4560.421 / 0.393
lag 10 ns4.88 / 3.982.59 / 2.280.381 / 0.3310.322 / 0.281
sampling change4.75 / 3.892.15 / 2.170.229 / 0.2480.242 / 0.257
348 K4.01 / 3.301.83 / 1.650.188 / 0.1590.196 / 0.178
lag 10 ns + sampling change + 348 K3.83 / 3.231.61 / 1.450.146 / 0.1280.181 / 0.168
paper's finetuning (300 epochs)3.73 / 3.171.56 / 1.400.136 / 0.1180.180 / 0.168

Q3

"The proposed representations are described as computationally efficient, but Table 1 reports approximately 91 hours of preprocessing for all mdCATH conformations, compared with 46 hours for the original Flood complex. Could the authors clarify the intended efficiency claim, particularly when the complete preprocessing cost is included?"

As we state next to Table 1, we trade some of the computational efficiency of the original Flood complex for a means to encode strucutral information. This makes mFlood around 2x slower than Flood. Nevertheless, we retain parts of the efficiency of the Flood complex, and mFlood is still approx. 3x faster than the Alpha complex (91 vs 300 hours).

The total of 91 hours is mainly caused by the large number of conformations (per temperature about 12.5 million: around 5,000 domains with 5 trajectories of about 500 frames each). Essentially, computing mFlood PH is a one-off preprocessing cost and parallelizes trivially across domains.

We are happy to have addressed several of the reviewer's concerns. As our previous reply to this comment was posted in the thread of reviewer g36h by mistake, we repeat it here.

Regarding the evidence for the generalization of our learned representations, we want to clarify that our updated manuscript establishes generalizability across protein sequences, dynamical regimes, and datasets:

  1. Our representations learned on the mdCATH training data capture the shape and dynamics of held-out domains, which are already chosen to minimize sequence similarity.
  2. These capabilities also carry into the generative setting, where our kinetic embedding trained on mdCATH induces ensembles closer to the reference on unseen fast folding proteins, whose dynamics are in stark contrast to those of mdCATH.
  3. We applied our representations to the ATLAS dataset, which is dominated by subtle dynamics. Again without training on this dataset, we have shown that our representations match conventional descriptors in the capture of slow dynamics and their induced Markovianity, and match or surpass them in the population of metastable states. Importantly, our representations identify these mentioned small-scale motions (Section 5.2, Appendix D.5, Figure 6).

To further strengthen our generalizability claim, the table below lists new results, several ensemble statistics for ATLAS test proteins as used in Jing et al. (NeurIPS 2024, Table 4), when the latter are generated with (i) MDGen, (ii) MarS-FM ft and (iii) our MarS-FM-mFlood ft. MDGen was trained directly on this dataset, whereas neither (ii) nor (iii) was.

Notably, our zero-shot MarS-FM-mFlood ft model closely matches (or even surpasses) the scores of the ATLAS-trained MDGen model.

ATLAS test (82 proteins)MarS-FM ftMarS-FM-mFlood ft (ours)MDGen
Trained on ATLASnonoyes
Pairwise RMSD r ↑0.450.520.48
Global RMSF r ↑0.420.490.50
Per-target RMSF r ↑0.650.680.71
RMWD ↓3.593.492.69
MD PCA W2 ↓1.401.411.89
PC sim > 0.5 ↑23%24%10%
Weak contacts J ↑0.470.480.51
Exposed residue J ↑0.500.520.29

The comparison with MarS-FM isolates the representation, as these models differ only in the features. MarS-FM relies on observables hand-picked for folding and unfolding. Our representation is tailored to no particular process, and yet, across all our generative experiments, it

  • matches or surpasses these observables on the unfolding dynamics of mdCATH,
  • surpasses them on the folding dynamics of the fast folding proteins, and
  • surpasses them on ATLAS, whose slow processes are local motions.

We also note that our aim is not state-of-the-art generation but to establish topological representations of protein structure and dynamics, for which the generative model serves as one downstream probe. In this light alone, the scope of our evaluation is already broader than published work. Established ML models of protein dynamics are usually trained and evaluated within the same dataset (e.g., AlphaFlow [1] and MDGen [2] on ATLAS, MarS-FM [3] on mdCATH), and transfer beyond it has so far been shown for small peptides [4, 5] or, in DeepJump [6], from mdCATH to fast folding proteins. Our representation, fitted once on mdCATH, transfers without training on either dataset to the fast folding proteins and to ATLAS.

Taken together, these experiments provide strong empirical evidence that our representations generalize well.

Does this summary, including the new comparison on ATLAS generative modeling, address your remaining concerns?

References

[1] Jing et al. AlphaFold Meets Flow Matching for Generating Protein Ensembles. ICML 2024.
[2] Jing et al. Generative Modeling of Molecular Dynamics Trajectories. NeurIPS 2024.
[3] Kapuśniak et al. MarS-FM: Generative Modeling of Molecular Dynamics via Markov State Models. ICLR 2026.
[4] Klein et al. Timewarp: Transferable Acceleration of Molecular Dynamics by Learning Time-Coarsened Dynamics. NeurIPS 2023.
[5] Daigavane et al. JAMUN: Bridging Smoothed Molecular Dynamics and Score-Based Learning for Conformational Ensemble Generation. NeurIPS 2025.
[6] Costa et al. Accelerating Protein Molecular Dynamics Simulation with DeepJump. NeurIPS 2025 AI for Science Workshop.

Reviewer_g36hOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper explores whether persistent homology can provide useful representations for protein structure and dynamics (without large scale prteraining). The authors introduce the masked Flood complex, which modifieds an existing topological construction to emphasize interactions between residues. They evaluate this in three settings: structural properties, kinetic embeddings, and improving conformational ensemble generation using MarS-FM.

Critical strengths

I like that the paper starts with a clear hypothesis that persistent homology could provide useful representations of protein structure and dynamics and then validates it from three angles (structural information, kinetic modeling, and downstream generative modeling). I think this makes for a well motivated study but some evaluations and proper methodological contributions could be strengthened/clarified.

Critical weaknesses

The main weaknesses are that the proposed method shows marginal/inconsistent improvements over the original Flood complex--obscuring the true contribution's utility. Also while the dynamics results are promising, the experiments and evaluation metrics are limited.

Questions for authors

marginal gains from the masked flood complex construction

The central contribution is presented as adding the protein specific masking operation to the existing flood complex. While i think this is reasonable, the improvements over regular Flood complex are relatively small and inconsistent across the benchmarks. The MSM results are more convincing, but i would like to see more clear demonstrations why masking is important and what additional information it captures

limited evaluation for dynamics

Most of the dynamics experiments are conducted on mdCath 450k where the slow process is protein unfolding. This is a reasonable starting point and the paper puts forth other uses, but to properly motivate this i would like to see whether these representations capture more subtle and biologically relevant motions: transitions between metastable states, allostering changes and pocket openings, also on other datasets like ATLAS or BPTI case studies. Would like to see generalization beyond high temperature unfolding.

metrics are limited for kinetic representations.

The improvements to MarS-FM are promising but are relatively modest and unfortunately are aggregate ensemble statistics like RMSF/RMSD. The main motivation is to learn representations capturing dynamics, I would have like to see more detailed distributional evaluations, such as PCA or tICA projections comparing the generated and reference ensembles.

7 comments

We thank you for raising targeted questions, the answers to which have led us to improve our submission, e.g., by substantiating the effects of masking.

Q1 marginal gains from the masked flood complex construction

"The central contribution is presented as adding the protein specific masking operation to the existing flood complex. While i think this is reasonable, the improvements over regular Flood complex are relatively small and inconsistent across the benchmarks. The MSM results are more convincing, but i would like to see more clear demonstrations why masking is important and what additional information it captures"

We agree that masking does not consistently improve performance across all probes. Its benefit mainly lies in kinetic consistency where, relative to Flood PH, masking lowers the deviation from Markovianity (CK RMSE) by a third and recovers the stationary reference distribution twice as well (π\pi-JS) at a slightly better VAMP2 score.

To lay out the effects that masking achieves, we have added section E.3 and G.2 to the manuscript, summarized below.

As both Flood and mFlood complexes are built from the same simplices, i.e., the Delaunay triangulation of the Cα landmarks, masking changes only the radius at which each simplex enters the filtration, i.e., its filtration value. For plain Flood complexes, a simplex is covered first by the atoms closest to it, usually those of its own residues. Its filtration value therefore strongly reflects the local covalent geometry, which changes little as the protein moves. Masking in mFlood excludes these atoms, and the simplex enters only once atoms of other residues reach it. Thereby, mFlood's filtration values report how closely surrounding residues pack around it.

In particular, as simplices corresponding to residues close in sequence cannot arise trivially, masking makes the filtration sensitive to secondary structure. For instance, an edge between consecutive Cα atoms enters the plain Flood complex at the same radius in helices (0.80 ±\pm 0.02 Å), strands (0.80 ±\pm 0.02 Å) and coils (0.80 ±\pm 0.04 Å) alike, set by bond geometry. Under masking, however, the filtration values of these same edges separate helices (2.86 ±\pm 0.15 Å) from strands (3.30 ±\pm 0.19 Å) almost perfectly (AUC 0.97 for mFlood versus 0.53 for Flood). Likewise, for triangles and tetrahedra of nearby residues (at most 4 residues apart), the separation of helices from coils rises from AUC 0.58 to 0.78 and from AUC 0.70 to 0.91, respectively.

Furthermore, as mFlood masks out the atoms of every residue that a simplex spans, but residues cannot simply reach inside one another, higher dimensional simplices receive systematically increased filtration values. This then results in an increase in the lifetimes of H1-features (loops) by a factor of around 3, and those of H2 features (voids) by a factor of around 7 (cf. Table 24).

A closer look at the cavities corresponding to the H2 features of the respective complexes reveals that most (72%) of the voids registered by masked Flood PH do not have a counterpart in Flood PH, as for the latter, they are filled exactly at the moment they are enclosed. In contrast, most voids that do have a positive lifetime under Flood PH live about 8 times longer under masking. Masking thus turns cavities that are invisible or short-lived into persistent H2 features. This holds specifically for the H2 features to which our kinetic embedding attributes the most importance (Appendix G.2, details in our answer to gPph).

The contribution of these features is highest in the folded state and drops as native contacts break (Figure 7,c-f). Because these features persist while the packing persists, mFlood vectorizations change slowly within a state and change when the structure rearranges, with about twice the autocorrelation of Flood PH (cf. Table 24).

Q2 limited evaluation for dynamics

"Most of the dynamics experiments are conducted on mdCath 450k where the slow process is protein unfolding. This is a reasonable starting point and the paper puts forth other uses, but to properly motivate this i would like to see whether these representations capture more subtle and biologically relevant motions: transitions between metastable states, allostering changes and pocket openings, also on other datasets like ATLAS or BPTI case studies. Would like to see generalization beyond high temperature unfolding."

We thank you for pointing us to further data sources. We have requested the BPTI data from D.E. Shaw Research but have as of yet not received an answer. Hence we can, at the moment, only present additional results based on the ATLAS dataset (Vander Meersche et al., Nucleic Acid Res, 2024).

To test generalization beyond unfolding, we applied our mFlood PH VAMPNet, trained on mdCATH at 348 K, without retraining to the 82 ATLAS test proteins (test split as in Jing et al. (NeurIPS, 2024)), which, simulated at 300 K, remain folded. Their slowest process corresponds to concrete motions (Fig. 6) such as the unwinding of a helix of the SSRP1 HMG domain and loop motions of the SARS-CoV-2 main protease. These transitions occur rarely, once or a few times per protein and not for every replica. As can be seen in the table below, our mFlood kinetic embeddings overall resolve as much slow dynamics as (Rg, SSF) or a TICA of SASA, and reproduce the stationary distribution of held-out replicas much better than the latter. They correlate with the individual subtle motions and, indeed, their slowest process identifies these motions in the first place (cf. Fig. 6).

MethodVAMP-2 ↑CK RMSE ↓π-JS ↓
mFlood PH1.189 ± 0.0090.043 ± 0.0020.123 ± 0.010
(Rg, SSF)1.175 ± 0.0100.043 ± 0.0020.125 ± 0.012
TICA (SASA)1.205 ± 0.0120.050 ± 0.0030.468 ± 0.009

Besides, in the generative setting, the paper already tests generalization beyond mdCATH on the twelve fast folding proteins with reversible transitions between folded and unfolded metastable states (see Section 5.3 ("Transfer to fast folding proteins") and Appendix D.4). There, mFlood-based representations improve upon the MarS-FM baseline in terms of (1) fidelity of covalent bond geometry, (2) folding free energy and (3) the distribution of secondary structure both with and without finetuning (cf. Table 23). They further reconstruct the kinetic distribution of the fast folder simulations better (cf. Table 22).

Q3 metrics are limited for kinetic representations.

"The improvements to MarS-FM are promising but are relatively modest and unfortunately are aggregate ensemble statistics like RMSF/RMSD. The main motivation is to learn representations capturing dynamics, I would have like to see more detailed distributional evaluations, such as PCA or tICA projections comparing the generated and reference ensembles."

We added both and now compare generated and reference ensembles in PCA and tICA projection space.

On mdCATH (see Table 18), we fit PCA on Cα distances and TICA on SASA of the MD replicas of each test domain and project the generated ensembles onto both. Our adaptation MarS-FM-mFlood is closer to MD than MarS-FM on all measures in both projections, with the largest gain in generating fewer conformations outside the MD distribution in PCA space (KL(gen‖MD) 2.69 vs 3.03, lower on 404 of 495 test domains).

We now also report similar results on the fast folders, where we project both generated and underlying MD ensembles via the reference TICA (see Table 22). Here, our mFlood adaptation is closer to the reference than the baseline MarS-FM on every measure, with and without finetuning.

Thank you for the response. The clarification and additional experiments address several of my concerns. However, I still find the overall empirical scope and evidence for the generalization of the learned representation somewhat limited. Therefore, I maintain my original rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.