Escribir un comentario

PathDiffusion: modeling protein folding pathway using evolution-guided diffusion

Kailong Zhao et al. doi: https://doi.org/10.64898/2026.01.16.699856

Reviewers: Alex Tong, Sofia Bali from James Fraser Lab @ UCSF

Zhao et al. present a diffusion model that generates folding pathways with interpretable states across the diffusion denoising schedule. The key distinction is using evolutionary guidance derived from sequence alignments to help convert between a purely sequence-conditioned model (Stage 1) and a sequence-conditioned+unconditioned model (Stage 2). PathDiffusion showcases exciting improvements in recapitulating “early” and “late” folding regions from experimentally supported pathways and shows general agreement with trajectories obtained from expensive molecular dynamics simulations. Pathdiffusion can sample locally unfolded states of both folded and intrinsically disordered structures derived from a sequence MSA to fold intermediate protein conformations. We found that the focus on continuous pathway generation presents a novel and compelling approach for understanding protein folding. The paper could be strengthened by providing more explicit definitions of success benchmarks (major point 1), clarifying model bias through potential data-set redundancies (major point 2), and addressing output reproducibility (major point 3), thereby improving its interpretability and applicability. Overall, the article presents needed advance in leveraging structural information in sequence conservation to bridge protein dynamics with their functions and model protein dynamics at a lower computational cost.

Major comments.

1. Broader comparison of folding intermediate success thresholds. The use of folding intermediate metrics, such as true positive conformations (TPC) and predicted pathway success, would be improved by providing the rationale and limitations of the selected cutoff values.

   • As the model focuses on point-specific noise schedules (PSNS), lower lDDT scores should, on average, be observed in regions with higher PSNS, even if the model has not learned the folding mechanics. Therefore, detailing the training dataset to show how the conservation and folding pathways are correlated can help clarify which improvements are from the model learning folding mechanics and which are expected from the clever addition of variability.

   • Regarding the folding trajectory performance metric: proteins with more experimentally labeled EFRs than LFRs would yield higher TPC ratios, but this would not reflect a true difference in prediction accuracy. Could this metric be improved by adding a protein-specific ratio that more broadly assesses the model's performance and accounts for the compositional diversity of the folded protein?

   • An interesting evaluation would be whether different protein classes or subclusters require different calibration for these thresholds.

   • More detail on the calculations would help model understanding, such as whether lDDT scores for early- or late-folded regions are averaged across the entire conformation?

2. The methodology used a 70% sequence identity threshold to cluster a curated PDB dataset of folded proteins. This threshold is higher than that of previous trained models, such as Alphafold2, which used a lower identity threshold of 40%.1 The higher identity restraint raises concerns of redundancy among homologous structures between training and evaluation targets. Including a detailed description of how the sequence identity threshold affects model performance would provide additional data on the model’s strengths and limitations.

   • Presenting model performance against the expected minimum TM scores for the high level of sequence identity would provide a clear view of the model's advances.

   • A more direct representation of how sequence redundancy affects model performance would be through evaluating success benchmarks on a filtered subset with stricter clustering. This would improve the interpretability of model performance, whether the output reflects a learned pathway versus interpolation across close homologs. This could also be synergistic with the ablations done at the end of the manuscript.

3. Pathway reproducibility and variation. The recapitulation of MD trajectories following the energy landscape in only 300 time steps is a very exciting prospect. One thing we found questions about was the reproducibility and sampling variability of the model. If different seed values were incorporated, would we see increased variability? Showcasing whether replicate runs sample consistent pathways or access different transition states would be an intriguing area of exploration.

   • In Figure 4, the comparison with BioEmu shows that the alternate model does not access the alternate states in a single trajectory. As the BioEmu model was trained on these MD simulations, it is necessary to know whether the final fine-tuned BioEmu model that includes the MD dataset was used and what parameters were used to run it.

   • For TICA, the collective variable used to build the energy landscapes was not described in the methods or text.

   • Are the pathways depicted throughout the manuscript from a single run of 300 time steps, and can these pathways be replicated? It would be interesting to see the pathways generated from single runs and compare the sampling landscape of this architecture with that of the BioEmu model chosen for the analysis. Another protein folding model (ref biRxiv) presents an interesting energy-based alternative for understanding folding pathways; having multiple runs for PathDiffusion could visualize this.

Minor comments.

1. Several description clarifications would improve audience accessibility. Since the central focus is an evolutionary guided model, it would help to include in the introduction a concise explanation of what MSAs encode and how conservation and covariation relate to structural constraints. Relatedly, it may be important to address how MSA quality factors such as depth, coverage, and gap fraction influence residual guidance in components of the model.

2. Typical protein engineering involves phi-value analysis derived from kinetic measurements to study folding intermediates and transition states. An interesting way to evaluate pathway hypotheses could be to define a proxy phi value. For example, comparing the probabilities of native contact formation of residues against an experimental phi-value profile can provide high confidence in predicted intermediate states. If not well matched, it could provide insight into where evolutionary guidance and kinetic observables diverge.

3. Lu, J., Zhong, B., Zhang, Z., & Tang, J. Str2str: A score-based framework for zero-shot protein conformation sampling. arXiv:.03117 (2023) This reference is given following the description of the TICA not included in the reference publication.

4. It was unclear to us whether Phi values are included in the FP52 benchmark. If not, this is a potentially rich source of information, as high phi values indicate contacts formed by the transition state (and, in this definition, early). This method could also resolve ambiguity about multiple folding pathways leading to intermediate phi values. This will also help us understand how deterministic this model is.

5. The TIM barrel result is very interesting. It presumably results from the differences in the dynamic weighting projection in 1D across the 3 TIM Barrels, leading to very different folding pathways as Stage 1 is converted to Stage 2? More explanation of what is driving the differences in the models would be quite helpful here.

6. This work relies heavily on chemical denaturant studies. The idea of differences between linear folding off the ribosome (Clark and others) vs. “all at once” out of denaturant should be discussed in a future direction.

7. The recent ProteinEBM could also be contrasted in the discussion; it shares many similar ideas. (https://www.biorxiv.org/content/10.64898/2025.12.09.693073v1)

[1] Jumper, J. et al. Highly Accurate Protein Structure Prediction with Alphafold. Nature 2021, 596 (7873), 583–589. https://doi.org/10.1038/s41586-021-03819-2.

Competing interests

The authors declare that they have no competing interests.

Use of Artificial Intelligence (AI)

The authors declare that they did not use generative AI to come up with new ideas for their review.