acceptodds
Under review as a conference paper at ICLR 2027

AnchorPose: Geometry-Aware MOF Assembly through Meso-Grained Pose Generation

Abstract

Predicting metal-organic framework (MOF) structures from given building blocks requires recovering their positions and orientations in a periodic crystal. The spatial effects of rotation errors are geometry-dependent and anisotropic: the same angular error can produce different atomic displacements depending on block size, shape, and rotation axis. Angular error alone, without reference to the specific block geometry, therefore cannot fully describe the spatial consequences of a pose error. We introduce AnchorPose, a meso-grained pose generation framework that explicitly incorporates building-block geometry. AnchorPose uses a small set of representative atoms as anchors, combines local shape with the current spatial state for pose prediction, and generates anchor positions with Bayesian Flow Networks. Rigid alignment with the known local geometry then recovers the poses of complete building blocks. This representation connects point-level spatial prediction with block-level structural constraints, bringing block geometry directly into generation while preserving intra-block structure without treating all atomic coordinates as assembly variables. Experiments on the MOF structure prediction benchmark show that AnchorPose achieves higher structure match rates than existing methods.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
metal-organic frameworks, crystal structure prediction, generative modeling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_kLH6OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper addresses the task of predicting a MOF structure (the lattice, fractional center, and rotation of each block) from building-block (number and molecular graphs). The authors build on MOF-BFN, retaining its lattice and periodic-center prediction, and changing its rotation representation from quaternions to four “anchor” atom coordinates per block. Rigid rotations are recovered by Kabsch alignment. This improves the loose match rate from 36.4% to 39.8% and the strict match rate from 9.86% to 12.36%.

Strengths

Weaknesses

The methodological novelty is limited, as the contribution primarily changes the orientation representation using established components. The choice of four anchors is heuristic; three non-collinear points suffice in the noiseless case but are not tested. Noise produces nonrigid configurations that must be projected back to rigid poses.

The improvement over the controlled baseline is modest, and the results depend on the configuration. A supplementary control favors Bingham rotation states under different readout and supervision choices.

Critical strengths

The approach is geometrically motivated: the same angular error causes different atomic displacements depending on block size and shape. Anchor coordinates incorporate this dependence but doen't preserve block geometry (requires projection).

Critical weaknesses
  1. The methodological novelty is limited. The method only tweaks the orientation representation. Further, there is no clear argument for why four anchors should be the best rotation representation. It even comes with downsides, like that noise produces nonrigid configurations that must be projected back to rigid poses.

  2. The improvement over the controlled baseline is minor. In Appendix C.6, Table 12 the anchors performs even worse then the Bingham (quaternion) rotation states.

Questions for authors
  1. Why are three anchors not tested? Three points should suffice (as long as they are non-collinear) in the noiseless case.
  2. Can you ablate the representation (between Bingham quaternions and Gaussian anchors) separate from the noise schedule? They use different native noise processes, but one could match spatial noise throughout the trajectory?
Reviewer_in3bOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses MOF structure generation from given molecular building blocks. Prior methods either represent each block in terms of its translation and rotation or directly generate all atomic coordinates. AnchorPose introduces an intermediate, meso-grained representation in which a small set of anchor atoms is generated for each block, with the corresponding rigid-block pose recovered via Kabsch alignment. The main ablation results indicate that using anchor coordinates as the generative state is more effective than using a rotation-space representation, whereas using anchor prediction only as the final readout yields relatively limited gains.

Critical strengths
  • The use of anchor points for pose recovery has precedent in other domains, but employing an anchor-coordinate generative state for MOF assembly is novel in this context and achieves the highest match rates among the compared methods.
  • The proposed representation is well motivated by the geometry of pose errors: the same angular error can induce substantially different Cartesian displacements depending on the block’s size, shape, and axis of rotation.
Critical weaknesses
  • The improvements among several key ablation variants are relatively modest, while the reported models appear to be based on a single training seed. Without repeated training runs, it is difficult to assess whether these differences are robust to training randomness.
  • While the experiments support the benefit of using a Cartesian point-based generative state over a rotation-space representation, they do not clearly isolate the specific benefit of the proposed sparse anchor representation itself (see Question 2).
Questions for authors
  • Could the authors evaluate the main ablation comparisons across multiple independent training seeds and report the resulting variability?
  • Could the authors provide a controlled comparison in which all atoms in each building block are used as anchors, i.e., the number of anchors is set equal to the number of atoms in the block? In this variant, the model would generate Cartesian coordinates for all atoms, after which the resulting point cloud would be rigidly fitted to the known local block geometry using the same Kabsch alignment procedure as in AnchorPose. This would enable a more direct comparison between the proposed sparse meso-grained representation and a dense all-atom Cartesian representation under the same rigid-recovery procedure.
  • [minor] Please fix the garbled text in Figure 1.
Reviewer_hoDyOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces AnchorPose, a MOF structure prediction model that represents building blocks using anchor points, providing a middle ground between coarse-grained and atom-level models. Specifically, the authors train two components: (1) a conformation-fitting model that refines RDKit-generated local conformations to better match the reference crystal geometry, and (2) a BFN assembly module that generates the lattice, block centroids, and anchor coordinates, from which rotations are recovered via Kabsch alignment. The results demonstrate substantial improvements in match rates over both coarse-grained and all-atom baselines, particularly under the loose matching criterion.

Critical strengths
  1. Well-motivated approach: The motivation for the anchor-based representation is compelling and well justified. The authors clearly explain how the same angular perturbation can result in different atomic displacements depending on the building block's geometry, and demonstrate in Figure 5 how their method mitigates this issue.
  2. Strong empirical performance: The proposed method outperforms existing baselines in structure match rates, particularly under the loose matching criterion.
Critical weaknesses
  1. Independent conformation fitting: The approach of refining building-block conformations independently before assembly seems questionable. Interactions between building blocks can influence their local geometries, but the conformation-fitting module does not account for these interactions. This limitation may help explain why AtomMOF-L achieves lower RMSD and a comparable five-candidate strict match rate, where finer structural accuracy becomes more important.
  2. Misleading architecture diagram: Although the paper is generally easy to follow, Figure 4 seems misleading because it does not explicitly illustrate the two inter-block pose-refinement updates. This gives the impression that the predicted rotations and translations are not jointly refined to account for interactions between building blocks. Explicitly showing this module would make the architecture clearer.

Minor presentation issues (non-critical): In Figure 1(b), the words "for" and "geometry" overlap in the subtitle. Also, the AtomMOF results in Table 1 are inconsistent with those presented in the original paper.

Questions for authors

Have the authors considered a post-assembly refinement stage or an alternative approach that allows building-block conformations to adapt to interactions with neighboring blocks?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.