A Graph Completion Method that Jointly Predicts Geometry and Topology Enables Effective Molecule Assembly

Rohan V. Koodli, Alexander S. Powers, Amiya Pandit, C. Im, Ron O. DrorarXiv 2025 · DOI · arXiv

TL;DR

EdGr is a novel spatial graph diffusion model that solves the "molecule assembly" problem by jointly predicting atom positions and chemical bonds. By explicitly coupling geometry and topology during the denoising process, EdGr can reconcile noisy initial fragment placements to reconstruct molecules with significantly higher accuracy than methods that treat these tasks separately.

What Problem Does This Paper Address?

In structure-based drug design, a common strategy is to identify small chemical "fragments" that fit into specific sub-pockets of a target protein. The challenge lies in molecule assembly: connecting these fragments with covalent bonds to form a single, high-affinity drug candidate.

This is difficult because the initial "guesses" for where these fragments sit in the pocket are often imprecise. Traditional machine learning approaches for "spatial graph completion" (adding missing edges to nodes with coordinates) usually fall into two camps:

  1. Topology-first: They treat the noisy coordinates as fixed and try to predict bonds.
  2. Geometry-first: They try to move the atoms to their "correct" spots first, then add bonds as a post-processing step.

The authors argue that these disjointed approaches fail because geometry and topology are fundamentally interdependent. A chemist knows that a bond can only form if atoms are close, but also that the requirement of forming a bond might "pull" fragments into a specific orientation.

Methods

The authors introduce EdGr (Edge-Graph), a spatial graph diffusion model designed to reason over geometry and topology simultaneously.

  • Joint Diffusion: Unlike standard diffusion models that only noise/denoise node positions, EdGr’s architecture ensures that predicted edge likelihoods ($n_{ij}$) directly influence node position updates ($x_i$) at every denoising step. Proximity suggests a bond, and a high-probability bond exerts a "force" that guides spatial movement.
  • Rigid-Body Noising: Since chemical fragments (like benzene rings) are internally stable, the model treats them as rigid bodies. During the forward diffusion process, noise is applied to the translation and rotation of the entire fragment rather than individual atoms.
  • Architecture: The model uses Equivariant Graph Convolutional Layers (EGCL) inspired by EGNNs. It employs two parallel MLPs to learn features for known intra-fragment edges and potential inter-fragment "missing" edges.
  • Self-Conditioning: To improve stability, the model can "see" its own predictions from the previous denoising timestep, providing better context for the current step.
  • Inference & Constraints: During the reverse process, the model uses the Kabsch algorithm to ensure fragments remain rigid while they move and rotate to find their optimal bonding partners.

Datasets

The researchers used a dataset of approximately 35,000 protein-ligand complexes sourced from the Protein Data Bank (PDB).

  • Preprocessing: They decomposed known drug-like molecules into rigid fragments (e.g., phenyl, methyl, ethyl groups).
  • Synthetic Noise: To simulate the uncertainty of real-world fragment discovery, they applied random independent rotations and Gaussian translational noise (at 1Å, 2Å, and 3Å standard deviations) to the fragments before asking the model to reassemble them.

Key Insights and Ideas

  • The "Aha" Moment: Explicit Coupling. The most significant insight is that edge weights shouldn't just be an output; they should be an internal mechanism for spatial refinement. By making the coordinate update function dependent on the predicted edge weights, the model allows "connectivity cues to guide spatial movements."
  • Preserving Chemical Identity: By treating fragments as rigid units during diffusion, the model avoids "melting" the fragments into a cloud of atoms, which is a common failure mode in all-atom generative models.
  • Robustness to Displacement: Because the model can move fragments to satisfy bonding requirements, it is remarkably robust. Even when fragments are started 3Å away from their true positions, the "pull" of the predicted bonds helps the model recover the correct structure.

Key Results

EdGr was compared against geometry-only models (like EDM and Equiformer) and topology-only models (like GCNs and Graph Attention Networks).

  • Superior Recovery: At 1Å of translational noise, EdGr achieved a Full Molecule Recovery (FMR) of 64%, nearly doubling the performance of the next best model, EDM (38%).
  • Precision and Recall: EdGr maintained high precision (~85%) and recall (~86%) in bond prediction, significantly outperforming GCNs (23% precision), which struggled because they could not adjust the noisy input coordinates.
  • Noise Tolerance: Even at a high noise level of 3Å, EdGr maintained a precision of 71%, whereas other models' performance degraded much more sharply.
  • Chemical Accuracy: Molecules reassembled by EdGr showed a high Tanimoto similarity (0.88) to the original ground-truth molecules, indicating that the predicted bonds were chemically meaningful.

Limitations and Future Work

  • Rigidity Assumption: Currently, the model assumes fragments are perfectly rigid. In reality, some fragments have internal degrees of freedom (rotatable bonds) that the model cannot yet account for.
  • Bond Types: The current version only predicts the existence of a bond (single bonds). It does not explicitly distinguish between single, double, or triple bonds during the assembly process.
  • Generalization: While focused on molecules, the authors note that this "joint geometry-topology" framework could be applied to other spatial graph problems, such as reconstructing neural circuits from noisy microscopy data or 3D scene reconstruction in computer vision.

Summary updated

Abstract

A common starting point for drug design is to find small chemical groups or "fragments" that form interactions with distinct subregions in a protein binding pocket. The subsequent challenge is to assemble these fragments into a molecule that has high affinity to the protein, by adding chemical bonds between atoms in different fragments. This "molecule assembly" task is particularly challenging because, initially, fragment positions are known only approximately. Prior methods for spatial graph completion-adding missing edges to a graph whose nodes have associated spatial coordinates-either treat node positions as fixed or adjust node positions before predicting edges. The fact that these methods treat geometry and topology prediction separately limits their ability to reconcile noisy geometries and plausible connectivities. To address this limitation, we introduce EdGr, a spatial graph diffusion model that reasons jointly over geometry and topology of molecules to simultaneously predict fragment positions and inter-fragment bonds. Importantly, predicted edge likelihoods directly influence node position updates during the diffusion denoising process, allowing connectivity cues to guide spatial movements, and vice versa. EdGr substantially outperforms previous methods on the molecule assembly task and maintains robust performance as noise levels increase. Beyond drug discovery, our approach of explicitly coupling geometry and topology prediction is broadly applicable to spatial graph completion problems, such as neural circuit reconstruction, 3D scene understanding, and sensor network design.
BibTeX
@misc{koodli2025graph,
  author = {Rohan V. Koodli and Alexander S. Powers and Amiya Pandit and C. Im and Ron O. Dror},
  title = {{A Graph Completion Method that Jointly Predicts Geometry and Topology Enables Effective Molecule Assembly}},
  year = {2025},
  eprint = {2505.21833},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.21833},
  howpublished = {arXiv},
}
Figure