acceptodds
Under review as a conference paper at ICLR 2027

Destabilize to Stabilize: Inference-Time Guidance for Generating More Stable Proteins and Complexes

Abstract

We aim to design more stable proteins by steering protein diffusion models such as BoltzGen. We hypothesize that generative models learn an implicit plausibility landscape around generated proteins and that this landscape can be used for structure destabilization as well as corrective stabilizing sequence mutations. We introduce RipZipGuide, a training-free workflow based on inference-time guidance of diffusion models. Our destabilization technique operates on collective variables and elicits structural failure modes through progressive destabilization steering of generated proteins. A candidate protein or protein-complex is progressively destabilized over multiple noising-denoising iterations, with noise levels dynamically adapted to the candidate and with repulsive bias accumulated across iterations to discourage returning to previously sampled conformations, resembling metadynamics. Failure modes become observable as local unfolding events in nearby conformational basins. These regions are then patched/stabilized through targeted denoising using another bias optimizing for hydrogen-bonding. Case studies on a toy-model and realistic proteins demonstrate the universality, reproducibility, and utility domains of the method. We expect the main gain of the method to be in the speed-up of sampling proteins with custom properties without requiring any (re)training or fine-tuning of the base-model.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein design, Diffusion model, Inference-time Guidance, Enhanced sampling, Drug discovery, Proteins, Biology

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_CEurOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces a new inference time method called RipZipGuide that samples from a “stability” tilted distribution of a base model. RipZipGuide operates by (1) gradually destabilizing the protein via gaussian noise, (2) discovers unstable links, and (3) modify these unstable links via mutation to have stronger bonds. They demonstrate the benefits of this method on both synthetic and real protein datasets.

Critical strengths
  1. The heuristic of gradual destabilization to identify weak links and then replacing those weak links with mutations that strength the structure is unique and interesting
  2. The use cases foor generating more stable proteins are important
Critical weaknesses
  1. Writing is very hard to follow, especially the Methods section. For instance, Line 215-233 and Algorithm 1 circularly depend on one another to describe a single algorithm. Further, the description of the algorithm is very textually dense, making it hard to understand the brass tacks of the algorithm with many terms and verbs used without rigorous definition (e.g. “denoise” is used but what does this mean rigorously? Also, please see questions 3,4,5)

  2. I appreciate the problem of developing more stable proteins that the authors attempt to tackle. However, the motivation for why the proposed method needs to be the way it is very unclear. Not only is RipZipGuide extremely complex and heuristic (many arbitrary choices of cutoffs and thresholds throughout Algorithm 1 and the Methods section), the necessity of this complexity is not well motivated. Why wouldn’t a much simpler algorithm work? For example, define a scalar quantity for “stability” and then use any off the shelf inference time sampler to sample from the base distribution tilted by this “stability.”

Questions for authors
  1. Is there a more rigorous or quantitative definition of stability? The reason I ask this is if there’s some way we can quantitatively compute “stability,” could we just leverage any off the shelf inference-time conditional sampling method for diffusion models (like BoltzGen) rather than the used heuristic of destabilization + patching
  2. What does “DPS” mean (used in subtitle of 3.1 without definition)?
  3. Diffusion models operate via a noise conditioned score function. After noising the protein structure in step 1 of Section 3.2, which noise level of the score network (sigma in Eq. 1) do you feed this noised protein structure into for denoising? The lowest noise level or from the beginning (aka highest noise level)?
  4. RE Line 208, the authors use this notion of “destabilization” and make the logical equivalency of thermal destabilization = gaussian noising. More realistically, isn’t destabilization better represented as raising the temperature, rather than some artificial gaussian noising process that does not correspond to anything physically meaningful?
  5. The authors refer to “accurate” unfolded pathways, what does "accurate" here mean?
4 comments
AuthorsOpenReview ↗

Response to Reviewer CEur Part 1

Thank you for highlighting potential improvements in the Methods and writing, and for suggesting possible simpler approaches to address the goal we had in the paper.

We have updated the Methods to avoid circular references and undefined terms. Additionally, just above the Discussion, we now describe how each stage of RipZipGuide contributes to the overall result. We plan to add more ablation studies closer to the Nov18 deadline.

Improvable presentation

We have merged Algo1 and the 1-2-3 list to avoid circular referencing and usage of undefined terms.

Why wouldn’t a much simpler algorithm work? For example, define a scalar quantity for “stability” and then use any off-the-shelf inference time sampler to sample from the base distribution tilted by this “stability.”

We added a brief note in the Introduction and more details at the end of Results on how each part of the workflow contributes and (in Supp Section C.1) on how hyperparameters were chosen. Qualitative arguments from the existing data for all existing steps are given below. We plan to conduct an ablation study to quantify the contribution of RipGuide in the workflow and report it closer to the Nov18 deadline.

  1. The H-bonding score (eq.7) is a scalar stability proxy, chosen for the locality of its effects and its universality against different perturbations.
  2. Sampling independent proteins from the tilted distribution is sub-optimal because denoising from full noise costs wall-time while the bias contributes only at late stages. We instead use noise-denoise cycles, noising only to a few times a residue size - enough for local mutations that preserve the fold while increasing stability.
  3. Removing RipGuide would discard the information about which regions are weakest, forcing the stabilizing bias to be applied uniformly. This dilutes it (100% of the protein vs the weakest 10%) and adds variability in regions that were already stable. It also does not help, since unfolding of the structurally weakest region can cost the function regardless of how stable the rest is. We plan to report this ablation by the Nov18 deadline.
  4. RipGuide therefore locates the weakest regions, concentrating the bias there while leaving ~90% of the protein untouched.
  5. The no-bias ZipGuide resampling in Table 1 is itself an ablation of the stabilizing bias: bias-free resampling around the coarse fold already gives ~60% of the effect on H-bond counts in weak regions. Table 1 also shows total H-bond count can stay nearly unchanged while stability rises (rucaparib binder), so a global measure is less effective against partial unfolding - again motivating localization via RipGuide.
  6. We agree the exact hyperparameter choices are heuristic; however, the orders of magnitude of the important ones such as σ_d are bio/physically motivated (Results + Section C.1).

A more rigorous or quantitative definition of stability?

Thank you for emphasizing this ambiguity.

This is partially addressed above. We added more discussion to address the particular choice of a stability proxy in the Introduction (H-bond counts, for its generality and locality) and the benefits of certain different parts of the proposed workflow (Intro and end of Results).

  1. We agree that there can be different definitions of stability that would in turn require different approaches to improving them. One can consider changes in pH or salt, for which surface charge distribution can be very important, or mechanical stability for high-pressure conditions, for which protein compactness may significantly matter. Unfolding pathways induced by all these perturbations would likely somewhat differ from unfolding during thermal denaturation. However, the aim of this project was to find a measure of stability that would be more broadly applicable for common industry issues, mainly focusing on storage costs. While given a more specific stability definition, one can devise more effective ways to improve it, H-bonding is among the most general correlates of stability, meaning it can help against a wide range of perturbations. This is why a score related to H-bond count was used as the objective.
  2. On using existing methods with a given stability definition/proxy (e.g. an H-bond count smoothed for differentiability) - see our points above on the performance benefits of denoising from only partially noised structures and of localizing the mutated region.

What does DPS mean?

Fixed it, thank you for pointing out this omission. It means "Diffusion Posterior Sampling". We removed it from the title and added a non-acronym definition near where it is cited in Section 3.1, above eq.4.

AuthorsOpenReview ↗

Response to Reviewer CEur Part 2

Which noise level is fed in?

Thank you for pointing out this lack of context.

The noise dynamically grows (starting from very low magnitudes) to keep up with the basin exploration. At a given RipGuide step, the next noise level is based on RMSD-s of a few (usually 10) of the latest-sampled disturbed structures from the original. We also highlight this in eq.6.; Supp Fig5 shows that any constant noise level does not allow "physical" exploration, while this dynamical scheme performs well.

  1. We use an intermediate, dynamically increasing noise level. This is now explained in more detail in point3 in Section 3.2. You can also find more details in Section C.1. We start with a very low level of noise, however, we must increase it to reach structures progressively further from the original structure. This progressive increase in the noise level should happen only after the region of the currently-explored basin (in conformation space) covered by the current level of noise fills-up. Thus, we dynamically tie (eq.6) the noise level to the RMSD-from-original of several (usually 10) latest structures.
  2. Such a complicated approach may look unnecessary, similar to other heuristics you noted above. However, the Toy-model shows (Supp Fig.5) that this dynamics allows finding nearby basins in the “physical” order: using a too-low constant noise level does not allow any nearby basins to be discovered, while too-high constant noising discovers all basins (within the noise level), but does so in random order, which violates the termination logic of RipGuide, which assumes that samples deviating further from the original are (mostly) sampled after all basins closer to the original are sampled and filled-up.

destabilization = Gaussian noising?

Thank you for pointing out this confusing point.

Gaussian noise on its own was not meant to be presented as analogous to thermal destabilization. Instead, it is part of the noise-denoise cycles used by RipGuide to generate progressively disturbed structures of the fixed protein sequence (Fig3(b,e)). We meant to communicate that these disturbed structures resemble a protein gradually unfolding from thermal destabilization.

The sentence at line 208 (since changed) referred to the whole metaD-like destabilization approach as thermal-like, not to the Gaussian noise alone. Gaussian noise of progressively increasing magnitude (see our answer to your question 3) is repeatedly applied to X_orig, and each noised structure is denoised using the base-model vector field plus a metaD-like bias repelling from all previously denoised structures:

noise X_orig by N(0,e1) → denoise repelling from {X_orig} → X_1 → noise X_orig by N(0,e2) → denoise repelling from {X_orig, X_1} → X_2 → etc.

We claim {X_orig, X_1, X_2, …} should resemble the sequence of states sampled by physics-based parallel-bias meta-dynamics. The bias independently stresses all native contacts, so we expect the weakest to break first - which is RipGuide's main goal.

What "accurate" unfolding pathways mean?

Thank you for pointing out this ambiguity.

It was meant to mean "a sequence of protein structures (X_1, X_2, X_3, … from our previous answer) that accurately mimics how the given protein would unfold under quasi-equilibrium heating from its original basin/conformation". We extended the discussion on this in Section 3.2.

  1. For a protein, an “accurate” unfolding “pathway” is an ensemble of physical unfolding trajectories under a specific disturbance, e.g. a temperature increase. In principle, this requires an exact specification of the disturbance, e.g. the exact schedule of how the temperature rises or how the salt concentration changes in time. First, we do not have dynamics when using AI-informed sampling (at least not without extra work to get it), so we can only model quasi-equilibrium processes. In our case, we aim to mimic quasi-equilibrium unfolding from a temperature increase. A quasi-equilibrium thermal unfolding started in a basin would spend some time exploring the initial basin, then “spill out” to a nearby basin (likely over the lowest barrier, but not necessarily), then the next basin, etc. This is exactly how meta-dynamics works if it is given a collective variable general enough to describe any feasible partially unfolded state. This is also how RipGuide is shown to work in Fig2c: correct basins are discovered and sampled in the correct (physically-expected) order before all basins are filled-up and random space-exploration begins. Specifically, a slow thermal disruption would approximately uniformly “stress” all native contacts in a protein, which is exactly what our parallel-bias does to all native contacts.
  2. Fig3(c,f) show the trajectories begin with thermal-like motions, then (Fig3(b,e)) local loop motions, then significant unfolding - the order we expect of a real unfolding. This is not proof that the process is fully physical, but it exhibits no clearly unphysical features.

Thank you for the response to my questions and concerns. While I can appreciate the motivation and the steps it took to arrive at ZipRipGuide, I am still struggling to see the necessity of the complexity/heuristics of this method over a simpler off the shelf sampler for diffusion models. For instance, the author's point about having to do a full denoising run with off the shelf samplers. This is not the case: one could noise a portion of the protein (that we want to stabilize), send this partially noised structure to an intermediate noise level (doesn't have to be the smallest or largest), and then conditionally sample a more stable version of only the portion we want to stabilize (similar to image inpainting). This main concern is in addition to other presentation concerns that make the paper (in its current form) difficult to justify accepting.

As a result, I will maintain my score.

Thank you for specifying your suggestion.

We believe the workflow you suggested in your last response is very close to what RipZipGuide does. If we understood correctly, there are 3 stages proposed:

noise a portion of the protein (that we want to stabilize),

We use RipGuide to localize which portion of the protein to noise. Simpler metrics such as pLDDT will not work as well (internal/external a-helices, disordered tails: see our answers to Q9tW). We just obtained results from an RipGuide ablation study (will add to the Supp soon) - disabling weak-region localization (i.e., removing the RipGuide step) breaks the stabilizing capability of ZipGuide, so RipGuide is an important step.

send this partially noised structure to an intermediate noise level (doesn't have to be the smallest or largest)

This is the 1st stage of ZipGuide: noise the structure to an intermediate level (ereseqe_{reseq} in the manuscript, Section 3.3.1)

conditionally sample a more stable version of only the portion we want to stabilize (similar to image inpainting).

This is close to the 2nd stage of ZipGuide (Section 3.3.2): denoise while applying the "stabilizing bias" to the identified weak region (which is mathematically equivalent to using DPS-sampling for conditional sampling). We are denoising the whole protein, but only the weak-region is allowed to mutate, so this is exactly "inpainting" of the weak-region part of the protein sequence. The whole protein needs to be denoised (from the intermediate noise level) because the geometrical arrangement/fold of the residues that do not mutate can still change residues geometrically close to the mutated region.

Reviewer_Q9tWOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes RipZipGuide, a training-free, inference-time workflow for making generated proteins more stable by steering a pretrained diffusion model such as BoltzGen. RipGuide repeatedly re-noises a fixed starting structure and denoises it under a metadynamics-inspired repulsive bias accumulated over native-contact distances. ZipGuide then re-noises the original structure to roughly one amino-acid scale and denoises under a backbone hydrogen-bonding bias, letting BoltzGen’s co-design capability introduce stabilizing mutations in those regions. The method is benchmarked in a toy landscape and some protein systems.

Critical strengths
  • The core idea is interesting and novel
  • Incorporation and usage of collective variables and metadynamics for stable protein design is novel
Critical weaknesses
  • Algorithm 1 is not detailed and is very hard to understand.
  • The central validation is circular: patched structures are scored by survival under the same protocol, same base model, and same regions that defined the weak links, with H-bonds directly resisting the native-contact disruption that drives the blow-up criterion.
  • External stability oracles. Can the authors evaluate their claim of designing more stable proteins by evaluating stability with external oracles to evaluate the efficacy of the method, rather than measuring it entirely inside BoltzGen’s own landscape?
  • “Near the basin” and “partially unfolded” are not the same region. What sits immediately around a native basin is thermal breathing, loop motion, and rotamer exchange. Partial unfolding sits across barriers of several kcal/mol. The hypothesis equates the two by assertion. The cited conformational-sampling work does not bridge the gap either, since it recovers alternative folded states (apo/holo, open/closed) that are themselves well represented in the PDB.
  • The denoiser projects toward plausible-looking structures by construction, so trajectories that “visually resemble unfolding” say little about whether the landscape is right. Likewise, weak links landing in less-structured regions are a check that pLDDT alone would pass.
  • The number of weak links is set by hand rather than discovered. DBSCAN’s N_min is tuned so that roughly 10% of atoms end up in weak-link regions. A protein with one genuine weak spot and a protein with four receive the same treatment.
Questions for authors
  • Is there any validation of weak links against an external reference, HDX protection factors, φ-values, known unfolding intermediates, or MD-derived local unfolding?
  • What happens under no-bias resampling at compute matched to ZipGuide, measured by survival time rather than H-bond count?
  • How is the 10%-of-atoms threshold justified per system? What does the PN-score distribution look like for a protein with no genuine weak region?
  • For binding preservation, can you report interface RMSD, contact recovery, or an independent affinity predictor across more than one complex?
  • Can stability also be assessed under a destabilization protocol that does not reuse RipGuide’s own settings like different σ_d, p_blowup, a_metaD, or a different base model?

Some typos and stuff: “Botz2” (Figure 4 caption), “BotlzIF” → “BoltzIF” (Table 1 caption), “Figure 2b” is referenced as “the biased free-energy landscape” while the caption calls it “Filled-up landscape”

3 comments
AuthorsOpenReview ↗

Response to Reviewer Q9tW Part 1

Response 1/2

Thank you for all your feedback and suggestions of tests and independent verifications. We have updated Methods to improve the RipGuide description that used to be Algo1. We also added discussion of how destabilization events occur in the Toy-model and how this translates to physical-like (thermal -> loops -> unfolding) destabilization for real proteins. We describe how RipGuide destabilization and ZipGuide stabilization differ (to show that RipGuide verification is not fully circular) and why the remaining overlap balances stabilization effectiveness against generalizability. We plan to add more tests and independent verifications closer to the Nov18 deadline.

Algorithm 1 is not detailed and is very hard to understand.

We have merged the Algo1 into the 1-2-3 list to avoid duplicated information and circular references. The full algorithm for RipGuide is in Supp Section C.1.

The central validation is circular.

We added more detail in Section 4.3, above Fig.4: ZipGuide only uses H-bonds near the weak-region, while RipGuide stresses all native contacts in the whole protein, so the objectives are not exactly circular, although we recognize their overlap. Also, giving ZipGuide only limited information from RipGuide (e.g. weak-link regions but not the exact breaking contacts or breaking pathways) balances stabilization effectiveness and generalizability to other disturbances.

Some circularity is inherent: patching benefits from knowing how the protein responds to the disturbance. But too much of such information makes the patcher less generalizable. So, we give ZipGuide the weak region's location, but not which or how contacts break, to not overfit to RipGuide.

Can you evaluate stability with external oracles?

We plan to investigate stability using more independent methods and come back with results closer to the Nov18 deadline.

We initially stayed inside BoltzGen's landscape for interpretability: a structure BoltzGen calls a stable optimum may not be one for another model, so an external change is hard to attribute to the mutations rather than to the landscape.

"Near the basin" and "partially unfolded" are not the same region.

We now outline this at the end of the Introduction and add details in Results. Fig2c and Fig3(b,e) show how we see the basins' exploration proceeds from thermal-like motions to loop motions to unfolding, which also matches your expectations of how it should proceed:

  1. Fig3(c,f) confirm that changes sampled in the native basin have magnitudes consistent with thermal breathing, independent of the initial noise, and Fig3(b,e) show loop motions beginning shortly before significant unfolding: step 10 vs 50 in Fig3b, 80 vs 85 in Fig3e.
  2. We agree ConforMix work does not fully address our aim - it targets alternative folded states, in which not many native contacts are lost, so it's not a structural loss. For us, once the metaD-like bias fills the native basin, sampling is forced over the barriers into nearby, less stable basins: unlikely to be found by ConforMix, but exactly the partially-unfolded structures we need.

The denoiser produces plausible structures by construction; weak links in less-structured regions is a check pLDDT would pass.

We replaced the "visually resemble unfolding" with specific analogies with the Toy-model results and with expectations from physics. About pLDDT, we added explicit reasons why we believe PN-scores provide information not contained in pLDDT (which is quantified in Section C.2 and Supp Fig.14). Also added qualitative examples of the disordered tail and externally- vs internally-oriented residues in an a-helix that likely would be mislabeled by pLDDT but not PN-scores.

  1. We agree the whole method relies on the base model predicting structures outside of its immediate optima that are not only "plausibly-looking" but also still bio/physically-meaningful. There is literature suggesting this is often the case: ConforMix, and from a different angle: Roney, Ou & Ovchinnikov, Protein diffusion models as statistical potentials, bioRxiv 2025, 10.64898/2025.12.09.693073. Comparing to MD pathways is harder due to the mismatch between an MD force-field and BoltzGen's "force-field". Thus, we do the Toy-model study in Fig2 + Supp Fig5-12, where barrier-separated basins (as in proteins) are discovered and sampled in the physically expected order by a RipGuide driven only by a small diffusion model trained on Fig2a.

  2. On the pLDDT half: the correlation is real but weak, R ~ [-0.15; -0.7] across systems at both per-atom and per-residue resolution (Section C.2, Supp Fig14), so the two are not interchangeable.

  3. pLDDT vs PN-score

    • pLDDT would label disordered, floppy tails as weak regions, which is not useful
    • pLDDT would be high for both internally and externally oriented residues of an a-helix on a protein side, but only the former has a high PN-score (since the latter has no native contacts to lose)
AuthorsOpenReview ↗

Response to Reviewer Q9tW Part 2

Response 2/2

The number of weak links is set by hand rather than discovered.

In the last subsection of Section 4.2, we assume constant "per-residue density" on weak-regions for a given class of proteins, which explains choosing the number of weak-region atoms proportionally to the total number of atoms. The 10% was indeed heuristically chosen and is likely "overfit" towards the protein lengths and types used in our work. Systematically investigating different classes and sizes of proteins is an exciting direction we plan to pursue in future work.

On one weak spot vs four: RipGuide looks for the weakest region, not all of them. Several patchable regions are handled sequentially - fix the weakest, then whichever becomes weakest next. A region much weaker than the rest may stay weakest after patching, but it is then the bottleneck anyway, so patching the others would not help.

Validation against HDX, φ-values, intermediates, or MD?

Thank you for suggesting these independent verifications. This is partly answered below, as the RipGuide settings used for stability verification used a_metaD different from weak-region identification. We also plan to report more independent tests closer to the Nov18 deadline.

On HDX and φ-values: both need experimental data that does not exist for de novo designed binders, and φ-values also need a mutational series with folding kinetics. Applying RipZipGuide to an experimentally resolved protein (we tried the "vWF A1 domain in complex with caplacizumab": 7EOW pdb) proved more difficult in a way similar to how improving top1 out of large NinitN_{init} was not reliable. This suggests biologically occurring globular proteins are already often fine-tuned for stability sufficiently well so our simple H-bonding-only approach does not improve them further. A more extended ZipGuide-like approach addressing H-bonds + charge interactions + hydrophobicity might prove sufficient to stabilize such already-well-designed cases, which is an exciting future direction.

No-bias resampling at matched compute, measured by survival time?

Thank you for suggesting this additional control. We plan to come back with results closer to the Nov18 deadline.

How is 10% justified? PN-score distribution with no genuine weak region?

The end of Section 4.2 (pre-last paragraph) now discusses why we saw choosing the number of weak atoms proportional to the total atoms as approximately correct and what protein features (size, type) may affect this.

  1. 10% balances locality of the identified region against leaving ZipGuide enough "room" (sequence) to mutate. It is likely overfit to near-globular proteins of 100-200 residues; for much larger proteins, we would also cap the size of any single weak-region. Systematic study of protein classes is an exciting future direction.

  2. From our observations, PN-scores behave less reproducibly for proteins that have less-clear "concentration of instability". This is quantified by higher noise/signal ratios for such proteins in Fig13: Fig13(c,d) correspond to proteins that were not successfully improved by the RipZipGuide workflow, so they are more likely to have no clear weak regions. The noise/signal ratios for these are ~0.45, against ~0.2 or lower (Fig13(a,b)) for proteins that did improve with RipZipGuide. PN-scores(atom-index) look more uniform and show no extended regions with PN-scores significantly higher than average (Fig13(c,d) vs (a,b)), which quantifies "no genuine weak region".

Interface RMSD, contact recovery, or an independent affinity predictor?

Thank you for suggesting these additional tests of binding interface preservation. We plan to run additional investigations and come back with results closer to the Nov18 deadline.

Can stability be assessed without reusing identical RipGuide settings?

RipGuide used for stability verification used a_metaD=5, actually different from a_metaD=10 used for the initial run to find weak-regions. We mistakenly reported this as "identical" settings, for the reasons described below. This is fixed now. The end of Section C.1 also lists ranges of other RipGuide parameters we found to work best, so yes, these parameters can have certain variability without significantly affecting "physical" outputs of the method.

The red dotted line in Fig4d is the deposition count scaled by 10/5=2, using the ~1/a_metaD scaling confirmed in the Toy-model (point 5 of Section 4.1, Supp Fig10). This scaling collapses runs at different a_metaD, for the toy model and for the real mzb1-binder, which is why we loosely called the two runs "identical" - they only differed according to a pre-identified scaling. Varying sigma_d or p_blowup changes destabilization less predictably (Supp Fig5-12). A different base model risks observing an instability caused by a mismatch of what BoltzGen vs that-model call stable.

Typos: "Botz2", "BotlzIF", "Figure 2b".

Now fixed; thank you for pointing these out.

AuthorsOpenReview ↗

Updates with new data

We wanted to note that we added additional data and figures to address some of your comments:

Can you evaluate stability with external oracles?

We added a test with an external model TemStaPro (Fig.17) that generally confirms the stability trends observed from # of RipGuide steps until blow-up and from H-bonds counts

The central validation is circular.

Fig.17 addresses that. We also added Section D, which shows that proteins driven to conformations that are expected to be more stable (we biased compactness) indeed require more RipGuide steps to blow-up, which can be seen as independent confirmation of RipGuide as a stability measure.

No-bias resampling at matched compute, measured by survival time?

We added a new column in Table1 that addresses that. The trend is consistent with H-bonding-counts proxy: no-bias resampling helps somewhat, and bias-resampling helps more.

pLDDT would pass

In addition to qualitative examples of cases when pLDDT will label weak-links incorrectly, we now added Fig.18(a,b) that quantifies measures not directly present in pLDDT: high PN-scores regions are more buried and less H-bonding-satisfied than the rest of the protein, which would not necessarily correlate with pLDDT. And out Fig.14 indeed shows only limited correlation.

Validation against HDX

We added Fig.18(c) that compares PN-scores to a simple model for HDX-protection. There is no reproducible correlation between proteins. As the caption explains in more detail, PN-score is positively correlated with HDX due to # of contacts (how buried it is) but negatively correlated due to # of H-bonds (how H-bond-satisfied it is). Thus, different correlations may take over in different regions of different proteins.

Interface RMSD, contact recovery

We added Fig.19 that shows that both i-RMSD and f_nat are "acceptably" preserved (according to the threshold from [doi.org/10.1002/prot.26188]) for the MZB1-binder and well-preserved for Rucaparib-binder.

more than one complex

We added a row in Table1 that patches a Brilacidin binder successfully. IC50 is also preserved.

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

The authors propose a method to steer protein diffusion models like BoltzGen to generate more stable proteins. The method is inference time and does not require model retraining. It does this by first destabilizing the protein to elicit structural failure modes and then these regions are patched using successive denoising steps. This is done by first identifying the weak links and then clustering them using DBSCAN. It then does targeted optimization focused on hydrogen bonding to introduce stabilizing mutations in the weak link areas.

Critical strengths
  • The biggest strength of this approach is that this method operates entirely at inference time eliminating the need for expensive model retraining or fine-tuning.

  • The model seems to maintain its functionality (for example: target binding) during the operations.

  • The authors claim 100x sampling efficiency.

Critical weaknesses
  • The biggest drawback of this paper is its narrow focus on hydrogen bonding for stability.

  • In table 1, it seems lke the stability gains diminish when NinitN_{init} is large.

  • The paper has a distinct flavor of being written significantly with AI support and was difficult to follow.

Questions for authors

"Improving patched structures also proved less reliable, perhaps due to saturation of local stabilizing modifications. Thus, we believe that structures sampled by ZipGuide-BoltzGen are within distribution for base-BoltzGen, just skewed towards more H-bonded structures. "

I am not sure what this means.

4 comments
AuthorsOpenReview ↗

Response to Reviewer yNMv Part 1

Thank you for your questions and comments.

We now describe why we chose to focus on H-bonding (locality + universality). We also elaborate on why we interpret diminishing stabilization for large-NinitN_{init} cases as a sign of ZipGuide using the base-BoltzGen as guardrails to avoid unphysical structures. Finally, we updated the text to hopefully improve its presentation.

We plan to test S-S bonding for stabilization closer to the Nov18 deadline.

The biggest drawback of this paper is its narrow focus on hydrogen bonding for stability.

Thank you for pointing out this potential improvement.

TL;DR: We now comment (Introduction) on why H-bonding was chosen as the main focus: its universality and locality. We also plan to address ZipGuide stabilization via S-S bond creation and come back with results closer to the Nov18 deadline.

What we mean by "universality" and "locality":

  1. Universal to many proteins, unlike e.g. S-S bonding, which can only help if a cysteine happens to have its SH group oriented in a way that would allow another close-by residue to mutate to cysteine (to make an S-S bridge) without violating any other protein constraints.
  2. Capable of local adjustments, unlike e.g., the compactness of the protein core, which can help stabilize many proteins but would require more global adjustments and is likely to affect the overall protein fold, which is against what we aim at because we want stabilizing mutations to stabilize locally while preserving the overall fold.

In table 1, it seems like the stability gains diminish when Ninit is large.

Thank you for highlighting this trend. Yes, that is the case.

TL;DR: We now comment on this at the end of Section 4.3. This diminishing is an expected feature (not a drawback) because it implies we are not sampling out of distribution of base BoltzGen, thus wet-lab verification of BoltzGen apply to our structures, and we can claim we sample a tilted BoltzGen distribution instead of states infeasible according to the base BoltzGen.

This is not necessarily a drawback: The fact that gains diminish at large Ninit means that given a sufficiently wide pool of initial candidates, we are able to find a candidate that we cannot further improve via our RipZipGuide method. This implies that ZipGuide-BoltzGen (BoltzGen with an additional guidance term that encourages H-bonding-optimal geometry near the weak regions) does not go far from the base-BoltzGen distribution, which would be the case if we were able to “improve” any design from the original BoltzGen. This inability of ZipGuide-BoltzGen to sample far out of the distribution of base BoltzGen suggests that all wet-lab verifications of the base BoltzGen (which characterize base-BoltzGen distribution) are largely applicable to structures generated by ZipGuide-BoltzGen.

The paper has a distinct flavor of being written significantly with AI support and was difficult to follow.

Thank you for pointing out this AI flavor and potential improvements in the writing.

TL;DR: We have made several modifications (see the updated PDF) that may make the paper easier to understand now

  1. List of modifications that may improve presentation:
    • Merged Algo1 with the 1-2-3 list that described it to avoid circular referencing and using undefined operations
    • Restructured the Introduction.
    • Listed how each part of the RipZipGuide workflow contributes to the end result
    • Expanded on why we focus on H-bonding for stabilisation: its universality and locality
    • Outlined how “metaD-unfolding” events proceed and how it corresponds to our physical expectations of the unfolding
    • Extended explanation on how PN-scores provide information missed by pLDDT
    • Emphasized that despite RipGuide and ZipGuide partially mirroring approaches, they have important differences (all native contacts for RipGuide vs H-bonds only for ZipGuide + no info about unfolding pathways for ZipGuide)
    • Added biophysics-based estimates we used for hyperparameters, especially non-dimensionless ones.
    • Extended explanation on why a value near 10% was chosen for labeling top-10% of atoms (according to PN-scores) as weak regions
    • Expanded on why we conclude we sample within the distribution of the base-model from being able to improve only structures with low NinitN_{init}
  2. We wanted to emphasize here that AI was used, as stated in our AI disclosure, “to aid or polish writing, … for retrieval and discovery, … for research ideation or execution”. About “ideation and execution” - AI was very helpful for automating testing such as hyperparameter scans and setting up jobs on HPC clusters. It also helped to polish math derivations of guidance term scalings (in the Supp Section A). However, the detailed architecture of RipGuide + ZipGuide, as well as additions to it that we found necessary during our work (e.g. the need for dynamically increasing noise levels in RipGuide) were conceived and designed by us.
AuthorsOpenReview ↗

Response to Reviewer yNMv Part 2

"Improving patched structures also proved less reliable, perhaps due to saturation of local stabilizing modifications. Thus, we believe that structures sampled by ZipGuide-BoltzGen are within distribution for base-BoltzGen, just skewed towards more H-bonded structures. " I am not sure what this means.

Thank you for pointing out this unclarity.

TL;DR: We have expanded Section 4.3 to clarify that applying RipZipGuide a second time to already patched structures does not result in systematic improvements and how this implies ZipGuide-BoltzGen stays within the (wet-lab-verified) distribution of BoltzGen.

This was meant to say that applying the same RipZipGuide workflow to structures that were already patched by this workflow once does not reliably improve them further for the cases we tested. This observation implies ZipGuide-BoltzGen is capable of reaching a limit in improvements it can make, constrained by the coarse fold structure. Such a limit should exist for a workflow that does not violate any bio/physical laws that govern how proteins are structured, so explicitly hitting this limit implies the model cannot (is not willing to) do “bad unphysical things” to formally improve the geometry for H-bonding-score, which is the desirable behavior as opposed to satisfying a too-strong bias by violating bio/physics learned by the base-model.

Thank you for your response. Apart from the concerns made initially and the way the paper is written, it is difficult to increase the score. I will maintain my score.

AuthorsOpenReview ↗

Updates with new data

Thank you for your response.

We wanted to note that we added additional data and figures to address some of your comments:

Focus on H-bonding only

Section D (STABILITY FROM COMPACTNESS) now gives a simple example of pure DPS-biased sampling to get more and less compact proteins that differ in their stability. This is a much simpler way to obtain generally more stable proteins, however, this affects the global fold, while RipZipGuide aims to stabilize only via local patches to preserve the general fold.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.