acceptodds
Desk rejected at ICLR 2027

MotionPlan: Training-Free LLM-Guided Residue Motion Planning for Protein Dynamics Generation

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Protein dynamics generators can produce plausible states and temporally ordered rollouts, yet their objectives largely fit marginal distributions rather than the inter-residue coordination organizing transitions; coordination errors therefore accumulate over long horizons. Retraining for explicit residue-level control is costly and requires paired motion supervision rarely available in MD data. We introduce MotionPlan, a training-free framework in which a large language model (LLM) infers per-residue coordination roles and confidence from sequence and structural context. The plan becomes a localized, confidence-weighted intervention on a frozen generator's pair representations throughout reverse-SDE sampling, without parameter updates or coordinate editing. Because Pairformer-based generators share pair representations, this provides a reusable inference-time control interface. On the complete 82-protein ATLAS test split, MotionPlan reduces trajectory-level MAE from 4.60 Å to 3.309 Å on the stride-256, benchmark and improves frame/consecutive-frame Pearson correlation from 0.62/0.53 to 0.67/0.59. On the 100 ns benchmark, it lowers JSD from to and raises state recall from to while retaining comparable molecular quality. These results show that LLM reasoning can translate structural evidence into actionable motion priors for frozen protein dynamics generators.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein dynamics, training-free control, diffusion models, pair representations, inference-time intervention

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

MotionPlan proposes a training-free inference-time control mechanism for protein dynamic generations. The core idea is using LLMs to assign residues coarse coordination roles (low/rigid, medium/boundary, high/flexible) + confidence score. The residue labels are converted into signed pairwise coefficient then injected as additive offset to the frozen generator's pair representations at reverse-diffusion steps.

Critical strengths
  1. The problem formulation is useful, especially with the trends of applying LLMs for the biomolecule generations. MotionPlan identifies a deficiency of trajectory generators that matching conformational distribution doesn't necessarily imply correct correlated motion/transition kinetics. The proposed offset/intervention is lightweight but shown considerable help with the generation. The short-horizon ablations also indicate that shuffled assignments perform worse, and that repeated injection matters.
  2. The authors are trying to avoid obvious target leakage. The evaluator only receives conditioning-time info such as seq, SS, contacts, etc. while excluding future coordinates, RMSF, PCA/tICA coordinates, etc.
Critical weaknesses

Critical ones:

  1. My major concern is that the central latent-space intervention is not scientifically grounded enough. The pair-control direction is just single randomly selected unit vectors uu in the generator's dimensional latent space. I personally think there is no reason a priori that moving along + uu should represent "decoupling" while moving along -uu should represent rigid coordination. Pairformer generally don't have a physically canonical orientation (i.e., learned through deep learning representations, not physical features). In such "arbitrary" or neural representations (which are hard to explain), offsetting may have symmetric status before looking at downstream behavior. This issue is also mirrored in the report in the manuscript that replacing uu substantially degrades performance. So I wonder, why did the randomly chosen sign happen to correspond to the claimed physical semantics? If the vector was genuinely selected before any benchmark evaluation, this result is surprising and requires robustness analysis.
  2. The role semantics are LOW/MED/HIGH, and pair coefficients are constructed essentially from the mean of the two residue role values. I think this is a strong physical assumption. A residue being flexible doesn't imply that two flexible residues should be dynamically decoupled. Flexible loops can move highly coherently. Conversely, two locally rigid domains can carry large relative hinge motions. Thus, the method appears closer to a mobility prior than a genuine residue-residue coordination model.

Minor ones:

  1. Improved downstream generation does not demonstrate that the LLM has learned scientifically meaningful coordination. The LLMs receive features that are already strong handcrafted predictors of mobility (like SS, contact, graph centrality, packing, etc.). So I wonder does an LLM contribute anything beyond simple structural role. The manuscript includes "deterministic proxy evaluator", which performs worse than MotionPlan, but its definition is not detailed.
  2. Molecular-quality evaluation after relaxation may hide precisely the failures that matter. We can't know where the performance improvement comes from.
Questions for authors
  1. Why did the randomly chosen sign happen to correspond to the claimed physical semantics? Why should +uu correspond to decoupling and -uu to coordination in a latent space without a known physical axis? How was the random pair-space direction uu chosen?
  2. Does an LLM contribute anything beyond lightweight learned predictor operating on exactly the same descriptors?
Reviewer_TaBgOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MotionPlan is a training-free inference-time controller for frozen protein dynamics generators. An LLM assigns each residue a coordination role (LOW/MEDIUM/HIGH) and a five-level ordinal confidence; these are deterministically lifted into a signed pair offset and added to the pair representation at every denoising step. On ATLAS (82 proteins), the method reports improved short- and long-horizon dynamic metrics over ConfRover while preserving relaxed-structure quality.

Critical strengths

1. The problem framing is clear: pair representations are an underused control surface, and specifying inter-residue coordination at inference time is an interesting idea.

2. Using an LLM as a training-free optimizer for frozen MD generators is computationally appealing and conceptually interesting; the approach scales naturally with LLM capability, which gives it clear long-term potential.

3. Molecular-quality results (Table 3) provide evidence that the dynamic gains are not simply coming from strained or unrealistic structures.

Critical weaknesses

1. The pair-level semantics of the lift are not validated. The coefficient αt,ij=(si+sj)/2⋅(ρi+ρj)/2\alpha_{t,ij}=(s_i+s_j)/2\cdot(\rho_i+\rho_j)/2 is separable: it contains no pair-specific information, so any two pairs with the same role labels receive the same bias regardless of whether they are in contact, allosterically coupled, or spatially unrelated. The paper provides no comparison between OtO_t (projected along uu) and reference-MD-derived coupling (DCCM or contact-frequency maps). Table 4a only shows that the distribution of per-residue roles matters, not that the pair field is biologically correct.

2. No structural visualization or case study. Figure 3 reports trajectory-level statistics, not structures. There is no trajectory overlay, no contact-map evolution, and no highlighted residue-pair distance trace. ATLAS at 100 ns mainly captures local fluctuations, which is weak support for a claim about "dynamics." A temperature-stratified dataset (e.g., MDCATH) and a classic large-amplitude case (e.g., ADK open–closed transition) would substantially strengthen the dynamic claim.

3. Baselines are thin. STAR-MD is cited for the long-horizon protocol but not compared against, despite being the strongest ATLAS long-horizon baseline. There is also no non-LLM deterministic pair-bias baseline (e.g., contact-map-based), so the contribution of LLM role inference is not isolated.

4. Static LLM input vs. state-dependent dynamics. The LLM sees only the conditioning structure (Sec. 3.2). For ligand binding, allostery, or temperature-driven changes, a residue's role can invert with state. Appendix A compares only general-purpose LLMs; a science-trained backbone ablation is missing.

5. The locality factor uses sequence distance, not spatial distance, so sequence-adjacent but spatially distant residues receive strong weight. This is not discussed or ablated.

I like the idea of this paper, and the finding that a general LLM can help refine generated MD trajectories is interesting. However, my concern is that the empirical evidence is not yet solid enough to clarify the mechanisms, so I think the paper is not yet ready for publication.

Questions for authors

1. Can you report the correlation between the projected OtO_t and reference DCCM/contact-frequency maps for several test proteins?

2. Can you add MDCATH (temperature-stratified) and large-amplitude case study (e.g., ADK) with trajectory overlays and key distance/angle traces?

3. Can you add more recent baselines beyond ConfRover and MDGen, for example STAR-MD or other recent trajectory/ensemble generators on ATLAS, and a non-LLM deterministic pair-bias baseline to isolate the contribution of LLM role inference?

4. Does replanning during rollout improve over a static plan, and does a science-trained LLM backbone change role quality?

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

This paper proposes MotionPlan, a training-free framework that uses an LLM to infer residue-level coordination roles and confidence scores from protein sequence and structural context. These predictions are converted into signed pairwise offsets and injected into the pair representations of a frozen diffusion-based protein dynamics generator. Experiments on ATLAS demonstrate improvements in short- and long-horizon trajectory generation.

Critical strengths
  1. The paper addresses an interesting problem of improving temporal coordination in protein dynamics generation.

  2. The proposed approach is training-free and can be integrated into an existing diffusion model without parameter updates.

  3. The reported improvements across multiple evaluation metrics are promising.

Critical weaknesses
  1. Limited robustness analysis of the LLM evaluator. The method relies heavily on LLM-generated residue classifications and confidence scores, but their consistency and reliablity are not sufficiently investigated. In particular, the sensitivity to different prompts, repeated evaluations, and potentially incorrect classifications remains unclear.

  2. Insufficient justification for the latent-space intervention. The method uses a fixed randomly initialized direction to encode rigid and flexible coordination. It is unclear why this direction should correspond to meaningful physical behavior in the pretrained latent space. The lack of experiments across multiple random directions further raises concerns about robustness.

  3. Unclear necessity of LLM reasoning. The LLM essentially produces a coarse three-class residue classification based on structural descriptors. Although the paper includes a deterministic proxy baseline, it remains unclear whether an LLM is necessary or whether stronger conventional flexibility predictors could provide comparble results.

  4. Limited evidence of generalizability. The experiments focus on a single pretrained generator, while the paper claims broader applicability to Pairformer-based models. Moreover, the coordination plan remains fixed throughout generation, potentially limiting its effectiveness when proteins undergo substantial conformational changes.

Questions for authors
  1. How consistent are the LLM-generated residue roles and confidence scores across repeated evaluations and different prompts?

  2. How sensitive is the performance to the randomly initialized latent direction? Have the authors evaluated multiple random seeds?

  3. How does the method compare against established protein flexibility predictors or other non-LLM aproaches?

  4. Can the authors provide evidence that the improvements generalize to other pretrained dynamics generators?

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

Related papers

Discussion (0)

Sign in to comment.