acceptodds
Under review as a conference paper at ICLR 2027

Prediction Is Not Action: Reliability-Guided Analytic Refinement of Molecular Conformers

Abstract

Molecular conformer generation is an important task in computational chemistry and drug discovery. Despite substantial progress in generative models, generated conformers can still exhibit residual local geometric deviations in bond lengths and bond angles. Correcting these deviations depends on the geometric state of the current input conformer: predicting plausible internal geometry does not by itself determine the corresponding coordinate update. We therefore introduce a compact prediction-to-action framework for post-generation conformer refinement that explicitly separates local geometry prediction from input-conditioned corrective action. The resulting local correction signals are mapped to Cartesian atomic displacements using analytic derivatives of local molecular geometry, while a separate magnitude-control mechanism regulates the extent of the final intervention. Experiments show that our method substantially improves strict local geometric validity while inducing only limited changes to the input conformers. Furthermore, a model trained only on conformers from a single upstream generator transfers zero-shot to outputs from an unseen generator, yielding substantial improvements in local geometry without retraining. Together with its compact parameterization, these results support prediction-to-action refinement as a lightweight and transferable post-generation approach for correcting residual local geometric errors.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Molecular Conformer Generation, Conformer Refinement, Geometric Deep Learning, Molecular Machine Learning, 3D Molecular Geometry, Post-generation Refinement

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_G5ymOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a Prediction-to-Action framework for post-generation molecular conformer refinement. The key idea is to separate graph-conditioned prediction of preferred local bond and angle geometry from Source-conditioned corrective action. The correction stage combines learned reliability weighting, analytic geometric Jacobians, and adaptive update-magnitude control to refine molecular conformers while preserving their original structures. Experiments on GEOM-DRUGS demonstrate improvements in local geometric validity, including generalization to unseen molecular scaffolds and zero-shot transfer across conformer generators.

Critical strengths
  1. Interesting and practically relevant problem. The paper addresses an important problem in molecular conformer generation. This is a meaningful objective with potential applications in molecular modeling and drug discovery.

  2. Clear presentation. The paper is generally well organized and easy to follow.

  3. Promising empirical results. The experiments demonstrate substantial improvements in strict local geometric validity with relatively small structural displacement. The evaluations on unseen molecular scaffolds and zero-shot transfer to another conformer generator further support the practical usefulness of the method.

Critical weaknesses
  1. Insufficient motivation and validation for separating prediction and correction. The central methodological contribution is the explicit separation of preferred-geometry prediction and corrective action. Although the intuition behind this separation is understandable, the paper does not sufficiently explain why this design should be preferable to an end-to-end refinement model. An end-to-end model could potentially learn both the preferred local geometry and the corresponding corrective updates jointly. Furthermore, the current experiments do not directly demonstrate the benefits of this separation.

  2. Potential mismatch between predicted local geometry and the Source conformer. The prediction stage estimates preferred local geometry using only the molecular graph, without conditioning on the Source conformer. However, preferred bond lengths and angles may depend on the specific conformational state. So graph-only predictions may represent an average over different conformational states rather than the appropriate local geometry for the current Source conformer. It will be good to see if there are failure cases as an ablation study.

  3. Insufficient comparisons with simpler refinement methods. The current experiments do not sufficiently establish whether the proposed neural components and architectural complexity are necessary. For instance, end-to-end coordinate refinement and gradient-based local geometry optimization.

Questions for authors

See weakness

Reviewer_F9KfOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a lightweight model (0.77M parameters) that repairs residual bond-length and bond-angle defects in generated conformers, to bring them closer to physical molecular geometries. A graph-only network predicts ideal local geometry; a learned "Reliability" gate, the analytic bond/angle Jacobian and a learned step size τ turn the residuals into a single small Cartesian update. On ETFlow outputs, Validity3D rises from 38.6% to 59.3% with only 0.007 Å displacement, and the gain transfers to unseen scaffolds and to DiTMC. The problem is relevant and the experimental bookkeeping is careful. This sort of small correction is useful in many workflows as a clean-up step after generating a guess at the geometry of a conformer. Note that there are existing tools to do this task; historically this was often done by doing a few steps of optimization using a classical force field like CHARMM or a semi-empirical quantum chemistry method like GFN2-xTB, and I guess today some researchers would use MLIP. This reviewer is not sure about the pros and cons of using different clean-up methods.

Critical strengths
  1. The problem being addressed is important. Fine clean-up of generated or guessed molecular geometries is almost always needed.
  2. The existing methods for clean-up are not perfect, so development of new methods is desirable.
Critical weaknesses
  1. The corrections to the bond-lengths and bond-angles are based on the connectivity only, not the 3-d geometry. So any guess geometry is being adjusted towards the same bond-lengths and bond-angles. In contrast, established clean-up methods using force-fields or semi-empirical quantum chemistry methods are aware that the bond-lengths and bond-angles depend are affected to some extent by the dihedral angles, and so those existing methods, at least in principle, could be much more accurate for some conformers than the new method.
  2. I think it is unclear if the new method is actually faster than existing methods based on force fields. So possibly the new method is both slower and less accurate than some existing methods.
  3. The comparisons with the existing methods are not done in a convincing way, and many of the statements in the paper are inexact and over-sell the new method.
Questions for authors

Important question: Have you tried basing your method on a well-established force field instead of assuming harmonic oscillator independent of dihedral angles? I think it is possible your method could be faster or more reliable than conventional optimizations using the same force field, but I am not sure, it would have to be demonstrated.

Here is a long set of comments. These mix comments about clarity and suggestions for improvements that you might want to do in future work, assuming that the authors will want to improve this for publication, possibly in a different venue. I hope these are helpful: Review: "Prediction Is Not Action: Reliability-Guided Analytic Refinement of Molecular Conformers" (ICLR 2027 submission #38017) General comment The paper proposes a lightweight model (0.77M parameters) that repairs residual bond-length and bond-angle defects in generated conformers. A graph-only network predicts ideal local geometry; a learned "Reliability" gate, the analytic bond/angle Jacobian and a learned step size τ turn the residuals into a single small Cartesian update. On ETFlow outputs, Validity3D rises from 38.6% to 59.3% with only 0.007 Å displacement, and the gain transfers to unseen scaffolds and to DiTMC. The problem is relevant and the experimental bookkeeping is careful. I have three main concerns. First, the novelty is overstated: the method is one steepest-descent step on a harmonic bond/angle potential, the standard Wilson B-matrix machinery of geometry optimizers, with learned parameters. Second, the mechanistic claims are weakly supported: all action-control ablations move the structure less than the full model, and the authors' own routing study (Table 14) shows that conditioning the learned components on the Source, the paper's central idea, has no measurable effect. Third, the baselines are insufficient: no learned refiners, no displacement-matched classical baselines, no non-learned controls, no downstream task, and the V3D ceiling of the GEOM reference data is never reported. The paper could become a useful contribution if reframed more modestly and supported by these experiments. Since the requested changes go beyond presentation, I recommend major revision. Detailed questions

  1. Overstated novelty. (a) Appendix A.2 shows that the update is one steepest-descent step on a weighted harmonic bond/angle potential, using the transpose of the Wilson B-matrix. Each learned quantity has a classical counterpart: µ ↔ force-field equilibrium values, r/σ² ↔ force constants (e.g., the cited Lindh model), τ ↔ step length/trust radius. This is the machinery of redundant internal-coordinate optimizers (Gaussian's Berny optimizer, geomeTRIC). Please state this in the main text and position the work against these optimizers. (b) Separating a target from the step taken toward it is standard (line search, trust-region methods, Levenberg–Marquardt, inverse-variance weighting). The novelty is that these quantities are learned; the "first" claim (l. 076) should be softened. (c) The slogan "Prediction Is Not Action" claims superiority over two alternatives that are never tested: (i) action without prediction, an equivariant network (e.g., EGNN) that directly regresses per-atom displacements, at equal parameters and displacement; (ii) prediction as action, an exact back-transformation that sets bonds and angles to µ (dihedrals fixed). Ablations inside the framework cannot support the slogan.
  2. Ablations are confounded with displacement. All ablated variants in Table 3 move the Source less than the full model (0.004–0.007 vs. 0.010 Å), which alone can explain their lower V3D. Please compare variants at matched displacement (e.g., V3D vs. RMSD curves over τ), and explain how the fixed τ = 0.00432 Å was chosen.
  3. Source-conditioning is not supported by Table 14. With no learned component reading the Source (mask 0000), V3D is 56.9%, the same as the final configuration (0101: 56.8%); the authors' own estimated effects (+0.1–0.3 pp) are not significant. This contradicts the paper's central claim. What seems to help is a learned per-primitive weight, not Source conditioning. Please reconcile with Tables 3 and 13. With one seed per mask, Table 14 also cannot justify the final configuration.
  4. Inference-time ablations overstate component importance. σ is also an input to the Reliability and τ heads, so replacing it after training feeds them out-of-distribution inputs. Indeed, the inference-time swap costs 12.9 pp (Table 3), while retraining without learned σ (J0 vs. J1, Table 13) costs only 5.5 pp (albeit on a different cohort, with a single seed and a fixed step size). Please report retrained ablations (three seeds).
  5. The gain is essentially bond-length correction. Bond-only reaches 49.2% vs. 50.3% for the full model; Angle-only barely changes the angle error. Please state this and tone down the Bond–Angle allocation contribution.
  6. Single step. Why only one step with one τ per molecule? Since residuals are computable, iterating (and/or per-atom step sizes) may give much higher V3D at still small displacement. Please test it.
  7. Targets ignore the Source conformation. (a) µ is graph-only, so all conformers of a molecule are pulled toward the same values, although equilibrium angles and bonds depend on conformation (gauche vs. anti angles, twisted conjugated bonds, ring puckering). (b) The training Reference is a random conformer of the molecule, not one matched to the Source, so the model learns ensemble-average geometry. Please report how much each bond/angle varies across a molecule's GEOM ensemble relative to the V3D tolerance, and compare with training on the closest Reference conformer.
  8. Unfair classical baselines. MMFF94s and GFN2-xTB appear to have been run without restraints (1.0 and 0.41 Å RMSD), i.e. they also change the conformation. Please compare in the same regime: positional restraints at several strengths (V3D vs. RMSD curve), frozen torsions, and few optimizer steps.
  9. Missing non-learned controls. Ideal bond lengths are largely tabulated, the typed-context statistics already used for s_stat could also provide reference values, and the gain comes mostly from bonds (point 5). Please report: (i) the same update with µ, σ from the table, r = 1 and constant τ; (ii) exact correction to table values with fixed dihedrals; (iii) a few MMFF/UFF steps with only bond and angle terms. If these approach 59.3%, the network adds little.
  10. No learned baselines. Flow-Matching Refiner, GeoOpt-Net and GOLF are cited but not compared.
  11. V3D of the Reference data. Please report the V3D pass rate of the GEOM Reference conformers. Since xTB-optimized structures pass only 59%, the GEOM (xTB) references likely fail often too. Then the model imitates targets that V3D considers invalid, and part of the gain may come from pushing borderline cases just past the threshold.
  12. Practical relevance. Which application needs V3D-valid structures within 0.01 Å of the input? Please show a downstream benefit.
Reviewer_RNfZOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

In this work, the authors aim at addressing post-generation refinement of molecular conformers, which may be globally reasonable yet still carry localized bond-length and bond-angle defects. The authors claim that predicting what local geometry a molecular graph prefers does not by itself determine how to move a particular conformer's atoms, and factorize refinement accordingly. A Prediction stage reads only the graph and predicts a preferred value and residual scale for every bond and angle primitive. An Action stage reads the current conformer and turns those predictions into motion through a per-primitive reliability gate, analytic Jacobians mapping primitive corrections to Cartesian directions, and a molecule-level magnitude controller, with rigid modes projected out. The proposed method show improvement on 100-molecule ETFlow test set.

Critical strengths
  • The target problem is clear: addressing post-generation refinement of molecular conformers
  • The energy evidence is independent of the headline metric and makes metric-gaming implausible.
Critical weaknesses

The main concern is the experimental results do not fully support the claim of the proposed method. In particular,

  • No learning-based refinement baseline is compared against the proposed method.
  • The proposed method is mainly tested on ETFlow and DiTMC. But there are many other popular molecular conformer generative models that are not validated on. Besides, though the target problem is clear, it can be too focused on chemistry applications which may not be of great interest for broader ML community which ICLR is targeted at.
Questions for authors
  1. Can you compare against a learned refinement baseline, eg a Flow-Matching based refiner trained on the same source data?
  2. Can you report COV-R, COV-P, AMR-R and AMR-P on the benchmark? Since these are standard metrics people report on GEOM benchmarks.
  3. Is stereochemistry preserved? Bond lengths and angle cosines do not determine chirality, so a displacement built from their Jacobians is blind to it.
  4. What exactly does V3D check? Sec. 4.1 cites GenBench3D without thresholds; please give the criteria and the bond/angle, clash and planarity sub-components separately.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.