acceptodds
Withdrawn from ICLR 2027

WM-Diff: World Model-Guided Planning for Angular Diffusion in Protein Backbone Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Internal-coordinate diffusion provides a translation and rotation-invariant representation for protein backbone generation, but existing samplers remain largely myopic by optimizing only the current denoising state without considering future structural consequences. Due to serial forward kinematics, small angular perturbations can accumulate into large Cartesian deviations and reduce sampling stability. We introduce WM-Diff, a world model-guided planning framework that equips angular diffusion with lookahead capability through horizon-conditioned future prediction. WM-Diff learns future steric cost-to-go from deterministic reverse-diffusion rollouts of a frozen backbone generator rather than instantaneous structural penalties. During sampling, the world model provides differentiable future-cost feedback to steer trajectories toward lower downstream steric cost. On an unconditional protein backbone generation benchmark, WM-Diff improves self-consistency designability from 27.5% to 52.0% while reducing steric clashes from 15.0 to 1.8 without post-hoc structural relaxation. Mechanistic analyses show that future-cost prediction provides a more effective planning signal than instantaneous objectives while preserving structural diversity. These results demonstrate the effectiveness of world model-guided planning for protein backbone generation.

Primary area:
generative models
Keywords:
Generative Models, Diffusion Models, Protein Backbone Generation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_KrGROpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes using the current denoising result to predict the clash risk after K steps of denoising with a pre-trained world model, and then correcting the denoising direction. The world model is trained using the current denoising result and the true clash risk after K steps. The authors claim that this achieves a significant improvement in design and a decrease in clash probability.

Critical strengths
  1. The method proposed by the authors can be integrated into any diffusion model, demonstrating a degree of scalability.
Critical weaknesses
  1. This article has absolutely nothing to do with world models, yet it's named WM-DIFF, which is very confusing.
  2. This paper lacks originality; its method is very similar to "Derivative-Free Guidance in Continuous and Discrete Diffusion Models with Soft Value-based Decoding." Therefore, the method proposed in this paper is not novel.
  3. There is a significant discrepancy between Table 1 in the main text and Table 10 in the appendix; the weighted average in Table 10 is not equal to the result in Table 1. Furthermore, there is a discrepancy between the clash count in Table 1 and the clash count in Table 11, with 1.8 vs. 4.5. This is a very serious error.
  4. This article neglects to discuss the time and memory overhead of the introduced modules.
  5. This article does not discuss the reduction in diversity in detail. It arbitrarily selects the largest designable gamma and hastily concludes the ablation study.
  6. The authors avoid addressing the issue that a high beta sheet does not necessarily indicate good results, and fail to reveal the connection between this method and increasing the beta sheet.
Questions for authors
  1. The author needs to change the title.
  2. The author needs to conduct in-depth research on past work, at least citing the paper "Derivative-Free Guidance in Continuous and Discrete Diffusion Models with Soft Value-based Decoding."
  3. Please unify the data in tables 1, 10, and 11, and provide a reasonable explanation for the significant differences in the initial submission.
  4. Supplementary experiments regarding time cost and video memory cost
  5. Provide a reasonable explanation for the two issues: decreased protein diversity and increased beta sheet proportion after adding the module.
Reviewer_kv1BOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

WM-Diff augments FoldingDiff-style angular diffusion with a horizon-conditioned regressor Wψ(xt,t,k)W_\psi(x_t,t,k) trained to predict log steric cost after a kk-step deterministic reverse rollout of a frozen backbone. Its gradient w.r.t. xtx_t guides sampling, with gradient-norm clipping and a linearly decaying scale. On 782 unconditional backbones, designability and severe both improve.

Critical strengths
  1. The lever-arm motivation is real: under serial forward kinematics an instantaneous geometric penalty is plausibly myopic, and framing this as cost-to-go is a natural response.
  2. The ablations (horizon, guidance strength, schedule, clipping) are more systematic than most comparable submissions, and Appendix C validates the surrogate independently rather than resting on downstream metrics alone.
Critical weaknesses
  1. The novelty appears incremental. The method is classifier guidance with a different scalar: Eqs. (8)–(10) are the standard construction, and Section 1 concedes WψW_\psi learns no transition model and predicts no future structure, so neither “world model” nor “MPC” describes the mechanism. The sole departure, guiding on future rather than instantaneous cost, is never isolated, since Table 2(b) only compares against instantaneous analytical objectives. The missing arm is analytical EclashE_{\text{clash}} on the same kk-step rollout terminal state; if that matches 52%, WψW_\psi is merely an efficiency device.

  2. The main result is not attributed, and the decisive control is missing. Guidance optimizes a steric surrogate and Clash Count is a thresholded version of the same quantity, so 15.0 → 1.8 is largely constructive. The +24.5 points of designability is what needs explaining, and Table 2(b) undercuts the obvious story: clash-only guidance reaches 0.9 clashes (better than WM-Diff’s 1.8) at only 28.0% designability. So clash reduction is not the cause. The current comparison also confounds two variables, learned-vs-analytical and future-vs-instantaneous. Separating them require a third arm: analytical EclashE_{\text{clash}} computed on the kk-step rollout terminal state and backpropagated through it. Without it, the paper’s central claim cannot be attributed.

  3. Evaluation criteria and baselines lag the literature. scTM ≥ 0.5 is no longer the community standard (scRMSD < 2 Å with pLDDT has largely replaced it) and is especially permissive for the 210 chains ≤ 70 residues. There is also no novelty metric against CATH, so diversity, measured only by within-bin pairwise TM, cannot rule out guidance collapsing samples toward a few easily-designable training mode. Appendix B.4’s argument against comparing to frame-based models justifies excluding them from mechanistic attribution, not from the table; readers need to know what 52% means in absolute terms.

  4. Clipping appears always active, which disconnects several analyses from the method. In Table 12, Mean GtclipG^{\text{clip}}_t exactly equals τ\tau at every setting, implying clipping fires on ~100% of steps. If so, WψW_\psi’s output magnitude is discarded entirely and only direction enters sampling, in which case the calibration analysis of Appendix C.3 is not evidence for the method as deployed. It also means the τ\tau sweep (Table 12) and γ\gamma sweep (Table 11) probe the same effective step size and should be a single 2D analysis. The fraction of steps with ∥gtraw∥2>τ\|g^{\text{raw}}_t\|_2 > \tau is required, not supplementary.

Questions for authors
  1. Can you add an arm that computes analytical EclashE_{\text{clash}} on the kk-step rollout terminal state and backpropagates through it, to separate lookahead from learned-surrogate effects?
  2. Were 782 structures regenerated independently for each seed? If so, how is a 0.7% standard deviation across seeds reconciled with the ~1.8% binomial floor?
  3. Please reconcile the γ=0\gamma=0 clash count (15.0 vs 21.0) and the mixed γ=0.10\gamma=0.10 row in Table 11.
  4. What fraction of denoising steps have ∥gtraw∥2>τ\|g^{\text{raw}}_t\|_2 > \tau, broken down by tt and by chain length?
  5. Can you report designability under scRMSD < 2 Å, plus the distribution of max TM-score to the CATH training set?
Reviewer_gKtBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper takes on de novo protein backbone generation, specifically targeting two big headaches in internal-coordinate diffusion models like FoldingDiff:1. Short-sighted sampling: Current samplers only care about fixing the immediate denoising step, completely ignoring how today's angle tweak impacts the downstream structure down the line. 2. The lever-arm effect: Because forward kinematics builds link by link, a tiny nudge in an upstream angle gets blown way out of proportion down the chain. This causes downstream atoms to swing wildly, leading to bad steric clashes, unstable sampling, and poor designability. To tackle this, the authors introduce WM-Diff. It reframes reverse angle diffusion as a receding-horizon planning problem—pretty much like Model Predictive Control (MPC). The goal is to boost designability and cut down on atomic collisions, all without having to touch or retrain the base generative model.

Critical strengths
  1. It introduces world model / MPC-style look-ahead planning into internal-coordinate protein diffusion, expanding "guided diffusion" from an instant goal to predicting consequences over a time horizon.

  2. The diffusion backbone is completely frozen, and the world model is a lightweight external predictor. It requires zero changes to the generative model and no post-hoc relaxation.

Critical weaknesses
  1. It has only been tested on unconditional single-chain generation (up to 128 residues). Conditional generation, motif scaffolding, multi-chain, or function-constrained design haven't been explored yet, so generalization is still up in the air.

  2. The world model is just a residue-level steric proxy rather than a full molecular energy or dynamics model, making the clash proxy objective pretty rough (only checking Van der Waals overlaps).

  3. While beta-sheet content shoots up, the authors admit this doesn't directly prove better folding quality, so some of the causal conclusions are a bit shaky.

Questions for authors
  1. The paper trains WψW_\psi using MSE to predict the log-steric cost after kk steps of rollout, but it doesn't report the actual prediction error of the world model itself. If the predictions have a systematic bias, won't gradient guidance push the sampling toward the "low-cost zones according to the world model" instead of actual low-cost zones (essentially hacking or exploiting the proxy model)?

  2. Real sampling is stochastic, but the world model is trained using deterministic rollouts. Won't the distribution gap between stochastic and deterministic trajectories cause out-of-distribution (OOD) errors for the world model during inference?

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

Related papers

Discussion (0)

Sign in to comment.