acceptodds
Under review as a conference paper at ICLR 2027

LatentSpring: Harmonic Sources and Physical Corrections for Molecular Generation

Abstract

Generating three-dimensional molecules from atomic composition requires learning both connectivity and geometry. We present LatentSpring, a flow-matching model that combines a harmonic-mixture source with learned geometric and physical corrections. The source samples atom positions along random auxiliary trees using element-dependent length scales, without requiring a chemical bond graph. Recovery training teaches the flow to restore perturbed geometries, while two small equivariant networks refine its predicted endpoints using reference structures and force feedback. On 64 unseen compositions, LatentSpring achieves a graph- and-force yield of 35.35% across five fits, compared with 8.75% for GAGA (Qu et al., 2026) and 8.15% for EDM (Hoogeboom et al., 2022), at a GFN2-xTB force- RMS threshold of 5 eV/ ˚ A. Adding local geometry checks gives a yield of 9.63%, compared with 4.43% for the initial force-corrected flow. This improvement holds in all three additional training repetitions, and the complete method outperforms a continuation control with matched backbone updates. The endpoint formulation also allows the trained correction networks to improve EDM and GAGA without retraining. Sampling requires only neural-network evaluations, with no energy queries or geometry optimization.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI4SCIENCE

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_Bay2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies 3D molecule generation conditioned on composition: the atom types are given, but the bonds are not. The authors use flow matching with a harmonic-mixture source distribution whose parameters depend on the elements. The harmonic terms are defined on random auxiliary trees over the atoms. These trees only serve to correlate the initial coordinates and are discarded afterwards, so the method does not need a molecular bond graph. On top of this, training includes a geometric recovery objective, and two small equivariant networks refine the predicted endpoints using geometric and force-based supervision. At sampling time the model only runs network evaluations, with no energy calls and no geometry optimization.

Critical strengths

1.The initialization does not need a known bond graph. Random auxiliary trees with element-dependent length scales give the starting coordinates some spatial structure, and since the trees are discarded afterwards, the final molecules are not forced into a tree topology. I also appreciated that the authors derive the normalized density of this source, so the construction is well defined.

2.Geometric and physical supervision is moved into training, so it comes for free at generation time. The recovery objective improves local structure and the correction network learns from force labels, yet sampling still only calls neural networks, with no potential energy model or geometry optimization in the loop. This sounds like a real practical advantage.

Critical weaknesses
  1. (IMPORTANT) The presentation needs work. The paper feels rough in places: the main figure (figure 1,3,4) conveys little, the experimental sections are crowded, and a lot of text goes into pre-empting possible objections. As a result it reads more like a log of experiments than a paper with a clear story.

2.The harmonic source on its own contributes little. Without the correction modules, it improves the joint pass rate over a Gaussian source by only 0.98 percentage points, and the confidence interval includes zero. With the corrections added, the gain becomes 4.59 points. My reading is that most of the improvement comes from the correction modules, and the paper does not yet show that the harmonic source gives a stable, significant benefit by itself.

3.The strict success rate is still low. 35.35% of samples have a valid chemical graph and residual forces within the threshold, but only 9.63% also pass the local geometry checks. Most outputs therefore still fall short of full quality, and the connectivity and local geometry problems remain open.

Questions for authors

See weakness 2&3 above.

Reviewer_N9cPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2(−1)
Soundness
2
Read review
Paper summary

This paper presents LatentSpring, a flow-matching model that generates 3D molecular structures from fixed atom types and counts without an input bond graph. It builds a harmonic-mixture source using random auxiliary trees and element-dependent length scales, then uses a self-conditioned flow to generate molecular coordinates. Geometric recovery training and two small networks correct the predicted endpoints using reference structures and force feedback. Sampling requires no energy queries or geometry optimization, and bonds are inferred from the final coordinates.

Critical strengths
  • The endpoint corrections can improve different generators without retraining. Within the shared EGNN architecture, the FM-trained correction networks raise geometry-and-force yield from 2.83% to 9.96% for EDM and from 3.03% to 10.45% for GAGA, with their weights and strengths unchanged.
Critical weaknesses
  • The method fixes atom types and counts but generates connectivity and geometry without a molecular graph. This is a valid generation task, but the paper does not clearly explain which molecular design needs this setting serves, or what advantage it offers over generating a graph with the same composition and then generating its conformers. Since this setting is the main reason for introducing a source that does not require known bonds, the lack of a clear use case weakens the motivation for the method and limits the meaning of its reported gains for molecular design.
  • The added value of the harmonic source is not clearly shown. Most of the improvement appears to come from the endpoint corrections. In Section 5.3, the harmonic source improves geometry-and-force yield by only 1.71 percentage points over the Gaussian source, with a 95% confidence interval that includes zero. This comparison also leaves out the self-conditioning and recovery training used in the full model. The paper does not measure how strongly the auxiliary tree shapes the final connectivity. These gaps make it hard to tell what the proposed source adds to the complete method, even though the correction networks show clear gains.
Questions for authors
  • The method generates connectivity and 3D geometry from fixed atom types and counts, without a molecular graph. Could the authors explain how this setting would be used in practice? Conformer generation supports docking and conformational ensemble generation, while de novo design can explore molecules with different compositions. What does fixing the composition while leaving connectivity open offer for molecular design? When would this be more useful than first generating a molecular graph with the same composition and then generating its conformers?
  • How strongly does the final connectivity depend on the auxiliary tree? Could the authors report the fraction of tree edges that become inferred bonds, and compare output connectivity diversity when fixing the tree and resampling only edge displacements versus resampling both?
  • Does the harmonic source help mainly through easier source–target pairing during training or through the initial arrangement of atoms? Could the authors compare pairing/alignment methods while keeping the source marginal fixed to assess the role of training-time pairing?
Reviewer_uaigOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

LatentSpring combines a random-tree harmonic source, self-conditioned flow matching, geometric recovery training, and learned endpoint corrections for composition-conditioned molecular generation. Experiments evaluate raw molecular structures and test whether the correction networks transfer to diffusion models.

Critical strengths

The correction networks improve EDM and GAGA without further training, making them useful beyond the proposed model.

Critical weaknesses
  1. EDM and GAGA achieve similar geometry-and-force yields with the same corrections. The added value of the full LatentSpring design remains unclear.
  2. Most outputs passing the graph-and-force criterion still fail geometry checks. Passing this criterion is therefore weak evidence of molecular quality.
Questions for authors

Which geometry checks fail most often for LatentSpring, and how does this compare with corrected EDM and GAGA?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.