- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes RayMorph, a novel framework trained on transition trajectories between protein pairs with undesirable and target properties. RayMorph can therefore redesign proteins without requiring the hand-crafted constraints and conditions used in previous protein design models. The authors evaluate their framework on redesigning proteins to have shorter N–C distances.
Critical strengths
- The idea is novel, and supporting diverse redesign objectives is an important problem. In practical protein design, we often encounter requirements that are difficult to incorporate into a design model. RayMorph offers a potential solution to this problem.
- The authors demonstrate their idea by redesigning proteins with shorter N–C distances, showing the effectiveness of RayMorph.
Critical weaknesses
- One concern is the validity of the intermediate states generated by the bridge between S₀ and S₁. It seems that their validity is not guaranteed. Could this affect training performance?
- The performance of RayMorph appears somewhat limited. The paper presents a sophisticated data pipeline for co-design, but the consistency between its generated sequences and backbones is substantially lower than that obtained with ProteinMPNN-designed sequences for the generated backbones. Would a structure-only backbone design approach followed by ProteinMPNN sequence design achieve better performance?
Questions for authors
See weakness, the paper overall is novel and interesting, my major concern is the performance of RayMorph, and I look forward to the authors’ response to this concern.
4 comments
Thank you for the detailed response and for the additional experiments. The new baselines and the structure-only ablation help clarify the contribution of RayMorph, and I appreciate the planned revisions to the functional-preservation claims.
On Weakness #1, I remain only partly convinced. Unlike standard diffusion/flow models that start from pure noise, RayMorph starts from a valid protein, and its bridge states also involve discrete split/delete edits. For example, a split produces a child that inherits its parent's coordinates. The response describes how the supervision is constructed, but it doesn't quantify how invalid intermediate states affect training. It doesn't show geometric statistics of the bridge states, or an ablation on the bridge design. I would encourage the authors to add such an analysis in the revision.
Overall, I appreciate the authors' efforts. After considering the rebuttal, I have decided to maintain my score.
Thanks for your question. Below we show a new analysis suggesting that the specific concern you outlined—that the child inherits the parent’s attributes—is not a driving factor of training quality.
We define a metric of an intermediate state’s validity as the average of the following two scores, already introduced in the paper:
- Bonds %: the fraction of consecutive Cα pairs (i, i+1) between 3.0 and 4.5 Å apart.
- Clash-free %: the fraction of residues with no other Cα within 3.5 Å, counting only residues at least 3 positions apart in the chain (|i − j| ≥ 3).
Our starting observation (see table below) is that the metric is near 100% at t=0 and t=1 but drops to approximately 82% for t=[0.4,0.6]. This supports the point you mentioned about intermediate states being invalid but also indicates that even at their worst, the intermediate states are only moderately invalid.
To address the concern of child-inheriting-the-parent driving this invalidity, we next sought to attribute intermediate-state invalidity to three causes:
- Linear interpolation between and
- Sigma knob of the Brownian bridge marginal
- Child placement on top of parent (reviewer’s concern)
We argue that #1 and #2 are validity compromises any plausible architecture will have to make. Specifically, #1 is the standard interpolation framework used very commonly in diffusion/flow models; e.g., FrameFlow[1] and FoldFlow[2] use it for coordinate changes. As for #2, it is a necessary component for stochastic inference to enable sampling diversity (when prior is a protein not noise) and one-to-many learning in a generative approach. Thus it is only with #3 that a RayMorph-specific design choice comes into play.
As an alternative to child-on-parent placement, we introduce ‘triangular insertion’: placing a newborn residue 3.8 Å (one Cα–Cα bond) from both its parent and the parent's right neighbour. This directly eliminates the penalty on both bonds and clashfree scores.
The table below illustrates an ablation on how the 3 factors affect the validity of the intermediate states. Across 400 subsampled protein-pairs in a training run, each cell reports the average of bonds% and clash-free% on the entire protein chain across time in the bridge interpolation between and . σ is the noise scale of the Brownian bridge; σ = 0 reduces it to linear interpolation. The current RayMorph setting is shown in bold.
| Child Placement | Inherit Parent | Inherit Parent | Triangular | Triangular |
|---|---|---|---|---|
| t | σ = 0.1 | σ = 0 | σ = 0.1 | σ = 0 |
| 0.0 | 99.8 | 99.8 | 99.8 | 99.8 |
| 0.1–0.3 | 87.3 | 93.4 | 87.6 | 93.7 |
| 0.4–0.6 | 82.0 | 86.7 | 82.6 | 87.4 |
| 0.7–0.9 | 89.9 | 93.5 | 90.3 | 93.9 |
| 1.0 | 99.9 | 99.9 | 99.9 | 99.9 |
Based on this analysis, we see that the two necessary design choices—straight line linear interpolation and sigma noise in the construction of brownian bridges—are the major drivers of intermediate state invalidity. We argue that the current design choice of having the child inherit the parent is not the main driver of input invalidity and will most likely not affect training performance. This is likely due to sigma noise already perturbs the coordinate locations such that the child rarely overlaps the parent directly.
| Factor | Effect |
|---|---|
| Straight-line path | About 13 percentage points (100 → 87.4) |
| σ = 0.1 noise | About 5 percentage points (87.4 → 82.6) |
| Child inherit parent vs. triangular | About 1 percentage point (82.6 → 82) |
References
[1] Yim et al. (2023). Fast protein backbone generation with SE(3) flow matching.
[2] Bose et al. (2024). SE(3)-Stochastic Flow Matching for Protein Backbone Generation.
Dear reviewer, thank you for engaging with us constructively. In our response above, we show through a new geometric analysis that the invalid intermediates are moderately invalid and their invalidity are mainly the byproduct of the bridge construction which is crucial for our stochastic sampling objective, one to many learning, and a structure to structure generative modelling. If you have further concerns, please let us know. We would appreciate for you to raise the score if we addressed your concern.
Weakness #1: “One concern is the validity of the intermediate states generated by the bridge between S₀ and S₁.” “Could this affect training performance?”
Thank you for raising the concern that the invalid intermediate states generated by the bridge could affect training performance. The intermediates are not guaranteed to be physically valid proteins. As is standard in diffusion/flow protein structure design models, the intermediates are normally invalid but the final target the model is trained towards is a valid protein. To mitigate the risk of poor training, the intermediate coordinates are supervised by the conditional drift toward assigned endpoint coordinates x₁, and amino-acid labels are trained to match a₁ (see Section 3.3 of the paper). We also incorporated auxiliary geometric losses that promote structural validity. (see Appendix B.3)
Weakness #2: “Would a structure-only backbone design approach followed by ProteinMPNN sequence design achieve better performance?”
We agree that RayMorph’s generated sequences do not match the quality of ProteinMPNN-designed sequences. Please see §2 of the Overall Response for the structure-only ablation addressing this concern.
If we have addressed your concerns, we hope you will consider raising your score.