acceptodds
Under review as a conference paper at ICLR 2027

SPINET: Sheaf Protein Inverse Folding Network

Abstract

Proteins change shape as they function, yet most inverse folding models predict amino acid sequences from a single, fixed backbone. A central challenge in protein engineering is to design proteins that undergo specific motions, which requires accounting for how their structures change over time. This motivates inverse protein folding conditioned on protein motion. We introduce SPINET, which predicts sequences from molecular dynamics trajectories. It uses cellular sheaves to represent residue interactions within each frame and recurrent units to integrate information across frames, then predicts all amino acids in a single pass. We evaluate SPINET on mdCATH and ATLAS, where it outperforms all evaluated static and ensemble baselines in sequence recovery. On mdCATH, it achieves 56.7% top-1 recovery, compared with 44.5% for the strongest static baseline and 40.7% for the strongest ensemble baseline. We also evaluate whether the predicted sequences are compatible with conformations sampled along the target trajectory. On mdCATH, they achieve a median TM-score of 0.760, and structural recovery favors target conformations over unrelated decoys for 99.5% of test domains.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
sheaf neural network, inverse folding, protein dynamics

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_DovyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The authors propose SPINET, a protein inverse folding model that leverages molecular dynamics trajectories. SPINET employs cellular sheaves to propagate spatial information between residues within each frame, followed by an RNN to aggregate temporal information for each residue across frames. On the mdCATH 320 K dataset, the authors report that SPINET achieves a higher sequence recovery rate and structural confidence score than the compared baselines.

Critical strengths

This work extends protein inverse folding from conditioning on a single backbone structure or a few discrete conformations to leveraging entire molecular dynamics trajectories, addressing a relatively underexplored problem setting. The overall model design is also clear and well-structured.

Critical weaknesses
  1. Comparisons with baselines are limited to 320 K. The main paper shows that SPINET outperforms PiFold, MapDiff, and DynamicMPNN only on the mdCATH test set at 320 K. No results for these baselines are provided at 348 K, 379 K, 413 K, or 450 K. Therefore, the results at 320 K do not establish that SPINET would consistently outperform these methods at other temperatures. Table 6 in Appendix D further raises concerns. The authors explicitly state that the main model suffered from overfitting in the temperature experiments. As a result, they used an earlier variant of SPINET for all five temperatures, removing the MLP after the RNN and the dropout before the RNN, reducing the number of channels from 16 to 8, and training for only 8 epochs at each temperature. This table does not include any baseline results. Moreover, neither the validation curves demonstrating the reported overfitting nor the test results of the final model at different temperatures are provided. I believe a fair comparison would require fixing the architecture and hyperparameters of the final SPINET model and training and evaluating it at each temperature under a consistent experimental protocol.

  2. The baselines are not trained on the same type of structural input. PiFold, GVP, ScFold, and MapDiff use relaxed structures from the RCSB corresponding to the mdCATH domains. These models are neither trained nor evaluated using the molecular dynamics coordinates employed by SPINET. Although the data splits and sequence labels are aligned, the structural information available to the models is not comparable. The Traj–MLP baseline in Table 1 provides particularly relevant evidence. This model simply averages node features over time and has only approximately 23K parameters, yet it already achieves a top-1 sequence recovery rate of 40.6%. This exceeds PiFold (34.3%), ScFold (39.6%), GVP (27.0%), and the retrained DynamicMPNN (37.6%), while being comparable to the fine-tuned DynamicMPNN (40.7%). The fact that trajectory-derived geometric features without graph-based modeling already outperform most baselines suggests that a substantial portion of the observed improvements may come from the additional input information rather than the sheaf-based architecture itself. I suggest sampling multiple frames from each trajectory and using these frames to train and evaluate the static inverse folding models. Both single-frame performance and multi-frame predictions obtained by averaging logits should be reported. This would help disentangle the benefits of using trajectory information from those introduced by the sheaf-based spatial modeling and RNN-based temporal aggregation.

  3. The conformational pair selection for DynamicMPNN also deviates from its original experimental setting. In the original paper, DynamicMPNN selects conformations with the largest RMSD or samples structurally diverse conformations using weights based on 1 − TM-score. In contrast, this paper first retains the 10 frames with the highest pairwise TM-scores and then selects the pair with the highest TM-score at test time, effectively choosing the two most structurally similar frames. Such a selection strategy provides very limited conformational diversity and may underestimate the performance of DynamicMPNN.

Questions for authors
  1. Could the authors provide a comparison between the SPINET presented in the main paper and the baselines across different temperatures?

  2. Could the authors explain why the sequence recovery rate at 450 K is substantially higher than that at 320 K in Table 6 of Appendix D?

  3. Could the authors include a multi-state inference baseline for ProteinMPNN, where predictions are made on the same frames used by SPINET and the logits are averaged across frames before computing sequence recovery? Additionally, ADFLIP (Yi et al., ICML 2025) is also capable of multi-state protein design, and I suggest including its results for a more comprehensive comparison.

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

This paper proposes SPINET, a protein inverse folding framework conditioned on molecular dynamics trajectories rather than individual backbone conformations. The architecture combines cellular sheaf-based spatial message passing with recurrent temporal aggregation, followed by additional sheaf message passing and residue-level amino acid prediction.

The method is evaluated on mdCATH and ATLAS against static and multi-conformation inverse folding baselines. On mdCATH, SPINET achieves 56.7% top-1 sequence recovery, compared with 44.5% for the strongest evaluated static baseline and 40.7% for the strongest ensemble baseline. The authors additionally assess structural compatibility using AlphaFold3 with target-conformation templates and unrelated decoy controls.

Critical strengths
  1. A meaningful extension of the inverse folding problem. Conditioning sequence prediction on molecular dynamics trajectories is an interesting direction that addresses the limitations of single-structure representations. The proposed spatial-temporal architecture is well motivated by the changing residue interactions along a trajectory.

  2. Strong sequence-recovery performance across two datasets. SPINET substantially improves native-sequence recovery over the evaluated baselines on mdCATH, with additional supporting results on ATLAS. The topology-separated evaluation and architectural ablations strengthen the empirical contribution.

Critical weaknesses
  1. The contribution of temporal dynamics is insufficiently isolated. Although the model processes trajectories sequentially, the experiments do not establish whether temporal ordering itself is necessary. Improvements could partly arise from access to more conformations rather than learned motion dynamics. Frame-order perturbations and stronger matched-input temporal baselines would be important for separating these explanations. Furthermore, the DynamicMPNN conformer-selection protocol may not adequately represent the trajectory's structural diversity.

  2. The structural validation does not establish motion-compatible sequence design independently of the target templates. The AlphaFold3 evaluation supplies representative target conformations as templates. Although the target-versus-decoy analysis is useful, it does not demonstrate that the sequences independently recover multiple target states or undergo the intended transitions. The structural advantage over DynamicMPNN is also small, and DynamicMPNN performs better on lDDT.

Questions for authors

How does SPINET perform when trajectory frames are randomly reordered or reversed, or when the same conformations are processed without temporal ordering? Could the authors also compare against a parameter-matched temporal graph model using the same frame budget?

Why does the DynamicMPNN configuration select a pool of highly similar conformations, given its intended multi-state setting? How sensitive is the comparison to selecting structurally diverse conformer pairs?

Can the authors clarify the extent to which target-template conditioning contributes to the structural recovery results? In particular, are template-free or independently conditioned comparisons available, together with paired uncertainty estimates for SPINET versus DynamicMPNN?

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

SPINET predicts amino acid sequences from molecular dynamics trajectories using sheaf message passing and recurrent temporal aggregation. It improves native-sequence recovery on mdCATH and ATLAS and evaluates structural compatibility with AF3 target and decoy templates.

Critical strengths
  1. The sequence-recovery gains are substantial: 56.7% on mdCATH versus 44.5% for MapDiff and 40.7% for fine-tuned DynamicMPNN . The CATH-topology split and additional ATLAS results strengthen this evidence.

  2. I like the target-versus-decoy structural evaluation rather than relying only on native-sequence recovery. The sheaf model also improves over the similarly sized TGNN ablation, supporting the value of its spatial representation.

Critical weaknesses
  1. The experiments do not isolate the value of temporal order: there is no shuffled-frame or pooled-sheaf control, while DynamicMPNN receives only two conformers in Appendix A.2. The gains may partly reflect access to more structural observations rather than learned motion.

  2. AF3 template-based recovery establishes structural compatibility, but does not show that designed sequences reproduce the target dynamics. The median TM-score is also close to DynamicMPNN's, 0.760 versus 0.756 from Table2. so the large recovery gain has limited demonstrated impact on the motion-design objective.

Questions for authors
  • How much recovery is retained with shuffled frames or order-invariant pooling while keeping the sheaf encoder and frame budget fixed?

  • Do simulations of the predicted sequences reproduce target conformational changes beyond static template compatibility?

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of author-reviewer discussions

Dear Reviewers and AC,

As the discussion period draws to a close, we thank you for your time and thoughtful feedback. Your comments prompted a closer examination of our results, several clarifications, and a few additional experiments that helped strengthen the manuscript.

We appreciate the strengths recognized across the reviews. Reviewer Yxcm highlighted the sequence-recovery gains, the topology-separated evaluation, and the target-versus-decoy structural analysis. Reviewer fWGq recognized trajectory-conditioned inverse folding as a meaningful extension of the problem and noted that the architecture is motivated by changing residue interactions. Reviewer Dovy highlighted the relatively underexplored setting of inverse folding from MD trajectories and the clear overall model design.

SPINET predicts sequences from molecular dynamics trajectories, opening a path toward modifying an existing trajectory and proposing mutant sequences compatible with the desired conformational behavior. We believe this makes an important contribution to inverse protein design by incorporating protein motion into sequence generation, a capability central to designing switches, molecular motors, and enzymes whose function depends on conformational changes. Below, we summarize the clarifications, observations, and additional experiments prompted by the reviewers’ comments:

(1) Shuffling frames caused only a small decrease in sequence recovery. We explored shuffling frames within the MD trajectory windows and found that sequence recovery decreased only marginally. We believe this is because the variation in residue geometry and contacts over time carries more information than the precise frame order. Our learned restriction maps let neighboring residues exchange information about this variation, helping the model relate changes at one residue to changes at its neighbors. This interpretation is also supported by our Temporal GNN ablation, where replacing sheaf message passing with standard graph message passing substantially reduced recovery.

(2) BioEmu provides evidence of ensemble compatibility without target templates. Independent MD simulations of the designed sequences were infeasible during the discussion period. We therefore generated BioEmu ensembles from designed and wild-type sequences for 44 common mdCATH test domains. SPINET had the lowest mean contact error and tied DynamicMPNN on long-range contact error. These results support the ability of SPINET designs to generate ensembles with wild-type-like residue interaction patterns.

(3) We examined AF3 recovery without target templates. We ran AlphaFold 3 without templates or multiple sequence alignments and found low recovery and confidence even for the wild-type sequences. The wild-type control achieved a median TM-score of 0.215 and pLDDT of 55.27. This helped clarify why template-free structural recovery via AF3 provides a limited basis for judging the designed sequences.

(4) We clarified conformer selection and expanded our DynamicMPNN comparisons. Our original description contained a typo. We accidentally stated that we maximized TM-score to select conformations for training DynamicMPNN. We actually selected structurally dissimilar conformations by maximizing 1 − TM-score and have corrected this description in the revised manuscript. We also evaluated larger conformer sets, which improved DynamicMPNN’s top-1 recovery to 55.4%, compared with SPINET’s 56.7%. This increased computational cost, with training taking approximately 11 minutes per epoch for DynamicMPNN versus 3 min 34 s for SPINET on an NVIDIA H200. We additionally evaluated the pretrained checkpoint released by the DynamicMPNN authors. It achieved 31.6% recovery on our mdCATH test set, compared with 37.6% after retraining the architecture on our mdCATH training set. These comparisons establish stronger baselines and help us assess how the training dataset and number of input conformations affect recovery and runtime.

(5) SPINET retains a sequence-recovery advantage against stronger baselines. We trained PiFold, ScFold, and ProteinMPNN on multiple MD frames and averaged their logits across frames at inference. We also added multi-state ADFLIP. These stronger comparisons improved baseline recovery, but SPINET still achieved the highest top-1 recovery at 320 K, reaching 56.7%, compared with 53.3% for ADFLIP, 50.2% for PiFold, 48.5% for ScFold, and 43.7% for ProteinMPNN. We clarified the training subsampling intervals and runtimes in the manuscript.

(6) We extended the baseline comparisons to 450 K. At the highest temperature available in mdCATH, SPINET achieved 70.0% top-1 recovery, compared with 62.4% for PiFold, 58.4% for ProteinMPNN, and 34.6% for two-conformer DynamicMPNN. We believe the greater motion at higher temperatures may expose more residue-specific backbone geometries and dihedral preferences, making sequence recovery easier.

The revised manuscript incorporates these experiments and clarifications. We believe the results strengthen the case for trajectory-conditioned inverse folding and show how learned sheaf message passing can support new approaches to protein sequence design. We would appreciate your considering these additions in your final assessments and revisiting your scores in light of our discussions.

Once again, we thank the reviewers and area chair for their time and engagement. Your constructive feedback helped us improve the manuscript and sharpen our understanding of the results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.