- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes LatentFold, a protein structure prediction framework that replaces discrete homolog generation with continuous latent MSA representations. The method uses a sequence-conditioned generator, probe-specific attention, and permutation-invariant statistical matching to learn MSA-like features from protein language model embeddings. Experiments on FoldBench and CASP15 report improvements over several discrete MSA-generation baselines, particularly after structure-supervised adaptation through a frozen folding backbone.
Critical strengths
The paper challenges the assumption that evolutionary information must be represented through discrete homologous sequences. The proposed continuous representation provides a differentiable interface to existing folding models and removes the need for target-specific MSA retrieval during inference.
The evaluation includes multiple sequence-depth settings, several protein language model sizes, and analyses of individual loss components.
Critical weaknesses
The comparison does not isolate the benefit of continuous MSA representations because the strongest LatentMSA(E2E) model receives structural supervision unavailable to the discrete generation baselines, and the methods use different pretrained representation models.
The proposed coupling evaluation relies on distances between projected residue embeddings rather than direct biological coupling measurements, making it difficult to distinguish genuine coevolutionary information from residue similarity and conservation patterns.
Questions for authors
Can the authors compare LatentFold against a simpler continuous MSA adapter trained with the same structural supervision and frozen folding backbone to isolate the contribution of the proposed architecture?
Does the generated representation improve agreement with experimentally observed residue contacts or established coevolution measures, rather than only matching the natural-MSA embedding space used during training?
2 comments
These analyses address important aspects of my concerns. I will maintain my current positive score.
Response to Reviewer xJZ1
Thank you for your careful review. Below is our step-by-step response to your comments.
Response to Q1
We added this comparison in Table 4. The adapter uses the same data interface, training data, and training steps, with offline and E2E variants evaluated through the same frozen Protenix backbone. LatentMSA improves overall lDDT from 0.739 to 0.793 offline and from 0.775 to 0.823 after E2E training; TM-score rises from 0.833 to 0.853 and from 0.870 to 0.885. The adapter architecture and training settings are in our anonymous repository.
Table 4. LatentMSA versus a simpler adapter on FoldBench. Oracle results; mean ± standard deviation across targets, shown as chain lDDT / TM-score.
Response to Q2
We agree that feature-space similarity alone does not establish recovery of biological coevolution. We therefore added a contact-ranking analysis using latentMSA, projected natural-MSA features, ESM features, and natural-MSA MI-APC scores.
Contacts are within-chain residue pairs with Cβ distance below 8 Å; we use Cα for glycine and other residues lacking Cβ. We rank pairs by negative L2 distance for feature representations and by MI-APC for the natural MSA, then report precision among the top L/5, L/2, and L pairs.
Table 5. Contact-ranking precision. Feature distances and MI-APC use different scoring functions, so this is an exploratory comparison.
LatentMSA has similar contact precision to projected natural-MSA features and higher precision than raw ESM features, though it remains below MI-APC. This supports its agreement with the learned feature space. However, it does not establish a direct correspondence between the latent features and the contact map.