acceptodds
Under review as a conference paper at ICLR 2027

Time-Dependent Directional Guidance for Physics-Informed Molecular Diffusion Models

Abstract

Diffusion models have become effective generators of 3D molecular structures, yet geometric plausibility alone does not ensure physical suitability for structure-based design. A common remedy is to add energy-based guidance during sampling. We argue that this naive intervention can behave like a forceful external correction rather than a compatible part of the reverse diffusion process, producing two critical failure modes: temporal misalignment, where a fixed physical signal ignores the coarse-to-fine denoising rhythm, and directional conflict, where raw energy gradients oppose chemically plausible generative motion. We propose TDGM, a Time-Dependent Directional Guidance Model that provides a plug-and-play, lightweight framework for aligning physical guidance with both denoising time and update direction. TDGM uses a temporal schedule to synchronize guidance strength with different generation stages, a directional projection to filter conflicting gradient components, and multi-scale physical priors to cover geometry, steric safety, pocket interaction, and global placement. On 100 CrossDock pockets with 10,000 generated molecules, TDGM obtains the best Vina docking score, QED, and Lipinski score among the compared methods, while MolPilot remains strongest on diversity and normalized SA. The results show that properly scheduled and directionally filtered physical guidance improves binding-oriented and drug-likeness endpoints while preserving explicit trade-offs across molecular statistics.

Primary area:
generative models
Keywords:
molecular diffusion, physics-informed generative models, temporal scheduling, directional projection

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

This paper introduces TDGM (Time Dependent Directional Guidance Model), a lightweight, retraining free, test time guidance framework for pocket conditioned 3D molecular diffusion. The work targets two well known failure modes arising from naive energy based guidance during sampling: temporal misalignment, in which static physical corrections interfere with early stage coarse scaffold assembly; and directional conflict, in which raw energy gradients push updates away from chemically realistic diffusion trajectories. TDGM employs binary temporal gating to activate physical constraints within a selected denoising window, decomposes physical losses into multi scale energy terms (geometry focused versus pocket attraction objectives), and applies directional projection to constrain raw gradients onto physically valid pocket interaction directions. Evaluated on 10 000 generated molecules across 100 CrossDock pockets, TDGM yields improved Vina docking scores, QED and Lipinski metrics relative to its base PMDM sampler. These benefits come with clear trade offs, including reduced molecular diversity and non trivial guidance induced structural perturbations.

Critical strengths

This submission offers valuable conceptual insights for inference time physical guidance within pocket conditioned 3D molecular diffusion. With rigorous empirical validation, the proposed ideas could benefit the broader structure based drug design (SBDD) community. • Formal decomposition of guidance failure modes: The authors systematically frame temporal misalignment and directional conflict as distinct failure mechanisms for naive energy gradient guidance. This formulation is well motivated and helps contextualize common problems observed in physics guided molecular diffusion. • Gradient directional projection as a test time vector filter: Unlike most existing guidance strategies that only adjust gradient magnitude via timestep dependent scalar weights, the directional projection module operates on gradient direction. It delivers a plug and play, inference only alternative to more compute intensive importance sampling or resampling based multi objective pipelines. • Comprehensive mechanistic diagnostics in SUPPLEMENTARY MATERIAL: The extensive ablation studies and trajectory analyses (including penetration severity profiles and runtime wall clock measurements) show thorough investigation of how guidance behaves across reverse diffusion timesteps. These mechanistic examinations go far beyond simple benchmark table reporting.

Critical weaknesses

While the conceptual design is promising, the current manuscript contains notable empirical gaps and unresolved causal ambiguities that weaken its core technical claims. Major revisions are required before this work can satisfy ICLR’s acceptance criteria:

  1. Causal ambiguity for the directional projection mechanism (Equation 11) As noted in SUPPLEMENTARY MATERIAL C, directional projection modifies both gradient direction and gradient norm. By mathematical definition, vector projection reduces the magnitude of the output vector (∥Πi(graw)∥≤∥graw∥\Vert \Pi_i(g_{\text{raw}})\Vert \le \Vert g_{\text{raw}}\Vert). It remains unclear whether observed improvements stem from filtering chemically invalid gradient directions (the paper’s central claim), or merely from reduced gradient amplitude, which inherently produces gentler corrections and fewer structural artefacts. Without an equal norm controlled ablation experiment, causal evidence supporting the directional projection mechanism remains incomplete.

  2. Missing evaluation on community standard 3D conformational benchmarks (GenBench3D) Supplementary geometry focused diagnostics (Tables 8, 9: Bond MSE, MMFF strain statistics) offer descriptive insights but are insufficient for fair cross method comparison against the wider SBDD literature. Since TDGM is designed to enforce chemically realistic geometry under physical guidance, evaluation under the standardized GenBench3D protocol — and reporting of the Validity3D metric — is essential. This would verify whether TDGM mitigates high strain 3D conformations native to diffusion based SBDD models, without relying on post hoc energy relaxation.

  3. Insufficient differentiation from prior time dependent guidance schemes Time varying guidance magnitude has been established in prior publications (e.g., KinetiDiff, MolPilot). The manuscript does not sufficiently contrast its binary discrete temporal gating (IiI_i) and multi scale energy partitioning against widely adopted smooth, monotonic λ(t)\lambda(t) scheduling strategies. In particular, the rationale for fully zeroing guidance strength in late stage timesteps (where local clash refinement is conventionally critical) needs clearer justification.

  4. Modest empirical gains, diversity trade off and absence of wet lab validation Performance improvements over the PMDM baseline are incremental. Gains in drug likeness metrics are accompanied by meaningful reduction in molecular diversity. Furthermore, all assessments rely purely on computational scoring; there is no wet experimental validation of predicted binding activity. The manuscript should moderate over optimistic claims and explicitly discuss these important practical limitations.

Questions for authors

Equal-norm ablation for directional projection: Add a controlled ablation where projected gradients Πi(graw)\Pi_i(g_{\text{raw}}) are rescaled to preserve the original norm of grawg_{\text{raw}}. If this controlled experiment cannot be practically implemented, substantially soften causal claims for directional projection throughout the main text.

GenBench3D benchmarking: Run TDGM under the official GenBench3D evaluation protocol and report Validity3D alongside complementary conformational-quality statistics. Demonstrate whether TDGM reduces strained 3D outputs without post-generation relaxation.

Clarify novelty of temporal-gating design: In the Methodology or Related Work section, explicitly contrast TDGM's binary multi-scale gating with existing monotonic λ(t)\lambda(t) scheduling from prior work. Provide clear justification for completely disabling guidance in late denoising stages, despite the importance of local clash correction.

Tone adjustment and limitation discussion: Revise manuscript text to cleanly separate pre-existing time-scheduling techniques from TDGM's original contributions. Explicitly highlight the diversity-reduction trade-off and discuss implications for practical SBDD workflows.

Reviewer_K56KOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TDGM, a test-time guidance layer applied on top of a frozen PMDM sampler for physics-guided, structure-based 3D molecule generation. The authors argue that attaching energy-based guidance naively to a diffusion sampler suffers from two failure modes: temporal misalignment, where a fixed guidance strength ignores the progression of reverse diffusion, and directional conflict, where raw energy gradients oppose chemically plausible generative updates. TDGM operates without retraining the base sampler. At each reverse step, PMDM produces a denoising proposal, which TDGM then refines. A temporal gate first activates a physical energy that evaluates multi-scale differentiable priors: bond-length regularization, intra-ligand and protein–ligand clash suppression, pocket attraction, and centroid alignment, grouped into a geometry term E_geom and an attraction term E_attr. The resulting gradient is projected onto a per-atom pocket-interaction direction and applied with a guidance scale and norm clipping. The corrected proposal y_j is then passed through an SE(3)-equivariant update block to obtain the accepted state. Experiments use 100 CrossDocked2020-derived pockets with 100 molecules per pocket. On the main benchmark, TDGM reports the best docking score, QED, and Lipinski score among the compared methods, while a competing baseline retains higher normalized SA and diversity.

Critical strengths
  • Useful framing of the question: The paper offers a useful framing of a difficult problem in 3D molecule generation by separating two questions that are often conflated: when guidance should act, and which gradient components should be applied.
  • Honest reporting of the experiments: The paper states that the 10,000 molecules are nested within 100 pockets rather than being independent target-level replicates, that the aggregate table does not demonstrate consistency across pockets or seeds, that the ablation does not support a universal scheduling claim, and that the projection diagnostic cannot isolate direction from realized update magnitude.
  • Well-constructed matched-state intervention experiment: The temporal analysis branches from saved intermediate states and uses a numerical-shadow branch to verify that the modified and unmodified samplers agree to within numerical tolerance when network outputs are held fixed. This cleanly separates the effect of the intervention from implementation noise.
  • Benchmarking hygiene: Wall-clock profiling uses warm-up runs, randomized arm order, and CUDA synchronization, and the authors explicitly warn that the per-update efficiency ratio should not be read as an end-to-end speedup, since all arms execute the same number of denoising steps.
Critical weaknesses
  • Inconsistently specified or underdescribed components: The “atom prior”, which according to Table 2, accounts for roughly 80% of the gain in Vina. It is treated as a separate ingredient in both the ablation and the projection diagnostic, yet it is never clearly defined, and its relationship to the proposed method and to the base sampler remains unclear.
  • Underspecified method components: Several elements required for reproduction are missing. The paper gives no functional forms for the five energy terms, the gate bounds i_min and i_max, the schedules γ_geom and γ_attr, or the λ coefficients.
  • Weak connection between scheduling and gains: The complete variants in the ablation are separated by a small margin, and uniform scheduling achieves the strongest Vina and Q+V results of the three. The ablation therefore does not appear strong enough to establish an advantage for time-dependent guidance. The supplementary experiments support temporal dependence and collision-correction efficiency, but do not establish improved final molecular quality.
  • Weak connection between directional projection and gains: The complete variants in the ablation are separated by a small margin, and uniform scheduling achieves the strongest Vina and Q+V results of the three. The ablation therefore does not appear strong enough to establish an advantage for time-dependent guidance. The supplementary experiments support temporal dependence and collision-correction efficiency, but do not establish improved final molecular quality.
  • No uncertainty quantification on the headline comparison: The reported Vina improvement over PMDM is marginal at 0.054, yet the main comparison provides no confidence interval and no paired per-pocket analysis. Reporting such an analysis, as the appendix already does elsewhere, would strengthen the claim.
  • Binding claims using single docking function: Binding-oriented evaluation relies exclusively on Vina, and the small improvement over PMDM has neither uncertainty estimates nor independent corroboration. Independent rescoring or complementary structural evaluation would strengthen the evaluation.
  • Missing naive energy guidance baseline: The paper's central premise is that naive energy guidance fails, yet this is never tested against the proposed method. Table 1 contains no "PMDM + unprojected, unscheduled energy guidance" row. Appendix C does include an unprojected arm, but all four arms there share a uniform schedule, so the arm being compared is TDGM without temporal scheduling rather than the production method, and the comparison uses 20 pockets with a single checkpoint seed. A naive-guidance arm evaluated under the main protocol against full TDGM is needed to support the paper's framing.
  • Missing post-hoc relaxation baseline: Post-hoc force-field relaxation of the base sampler’s outputs, including pocket-aware relaxation where appropriate, would provide a direct comparison for the feasibility improvements reported in the diagnostics table.
  • No code availability statement or anonymized repository is provided: The paper does not include a source for reproducibility.
Questions for authors
  1. What exactly is the "atom prior", and what explains the improvement beyond it? In Table 2, atom-prior guidance accounts for most of the improvement over unguided sampling. Could the authors define the atom prior precisely and provide a controlled ablation separating its contribution from those of the energy terms, temporal scheduling, and projection?
  2. What is the guidance interval in Table 2? Which time windows do "full early" and "full late" correspond to?
  3. The experiments do not appear to isolate whether applying different schedules to the geometry and pocket groups helps. Are there experiments that locate when collision correction is most effective for each energy group separately?
  4. How robust is the headline docking improvement? Given that the Vina difference between TDGM and PMDM is marginal, metrics such as seed-to-seed variance and per-pocket win rates would help solidify the claim.
Reviewer_yGNDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TDGM, a training-free, test-time physics-guided framework designed to address two challenges in molecular diffusion generation: temporal misalignment and directional conflict. The method integrates time-dependent guidance scheduling, directional projection, and multi-scale physical priors to incorporate physical constraints into the sampling process of a pretrained PMDM. Experimental results demonstrate that TDGM achieves competitive performance on metrics such as Vina, QED, and Lipinski. However, despite its emphasis on being training-free and plug-and-play, the framework has only been evaluated on PMDM. The lack of experiments across different backbone generative models leaves its generalizability insufficiently validated.

Critical strengths
  1. The paper clearly identifies two important challenges in physics-guided diffusion models: temporal misalignment and directional conflict.

  2. The proposed time-dependent directional guidance framework is well structured, integrating time-dependent guidance scheduling, directional projection, and multi-scale physical priors. Each component has a clearly defined role and is supported by some experimental evidence. The paper also provides intuitive case studies.

  3. The method is training-free and can be directly applied to the sampling process of a pretrained PMDM without requiring retraining.

  4. The method achieves competitive results in the main experiments, outperforming the compared baselines on metrics such as Vina, QED, and Lipinski.

Critical weaknesses
  1. Insufficient evaluation of generalizability and robustness. Although TDGM emphasizes its plug-and-play nature, it is only evaluated with PMDM. Its applicability and robustness across different diffusion backbones remain unverified.

  2. The quality of figure and table presentation needs improvement. Figure 1 contains small text and inconsistent visual styles. Some elements appear to be taken from screenshots, with inconsistent background colors, while several pocket and ligand illustrations appear potentially AI-generated. Overall, the figure does not meet my expectations for an ICLR poster-level presentation. Additionally, the title “Training Process of TDGM” in Figure 1(a) may cause confusion given the method’s training-free nature. Table 1 also appears insufficiently checked, as Vina, LogP, and MW lack arrows indicating the preferred direction of optimization.

  3. Insufficient validation of the core motivations and mechanisms. Although the appendices provide relevant experiments, the paper does not systematically quantify the frequency of temporal misalignment and directional conflict or their impact on final generation quality. The specific coarse-to-fine guidance strategy also lacks sufficient validation. Furthermore, the directional projection relies on a pocket-aware direction rather than the model’s learned generative update direction, leaving the mechanism by which it mitigates directional conflict without direct experimental evidence.

Questions for authors
  1. Appendix B.2 reports sampling times, but do the additional computational costs introduced by physical guidance justify the improvements in final molecular quality? Could the authors further analyze the quality–efficiency trade-off between TDGM and unguided PMDM?

  2. How do the results in Appendices A and C support the proposed coarse-to-fine denoising rhythm and directional conflict? Could the authors provide statistical analyses across sampling trajectories and directly quantify the directional conflict between physical gradients and PMDM updates, as well as the improvements achieved through projection?

  3. What is the rationale behind selecting the 50–80% guidance window in Appendix B? Could the authors provide a sensitivity analysis of imin⁡i_{\min} and imax⁡i_{\max}, along with the effects of different guidance windows on final metrics such as Vina and QED?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.