Generating Molecular Conformer Fields

Yuyang Wang, Ahmed A. A. Elhag, Navdeep Jaitly, Joshua M. Susskind, Miguel Ángel BautistaSubmission to ICLR 2024 · Rejected

TL;DR

Molecular Conformer Fields (MCF) treats 3D molecular conformations not as rigid geometric graphs or collections of torsional angles, but as continuous functions (fields) mapping atoms and bonds to coordinates in 3D space. By coupling normalized graph Laplacian positional encodings with a diffusion model over function space, MCF achieves state-of-the-art conformer generation on standard benchmarks without requiring hardcoded chemical priors like rotatable bond templates or explicit $SE(3)$-equivariant architectures.

What Problem Does This Paper Address?

Predicting the low-energy 3D shapes (conformers) that a molecule can adopt in physical space is central to drug discovery, computational chemistry, and protein-ligand interactions. However, searching this conformational landscape is challenging because the space of valid conformations grows exponentially with molecular size and the number of rotatable bonds.

Recent deep learning methods predominantly rely on heavy domain-specific inductive biases: they either model rotational torsion angles along chemical bonds (e.g., Torsional Diffusion) or enforce strict rotational/translational equivariance through specialized graph neural networks (e.g., GeoDiff). While physically motivated, these biases often rely on brittle, non-differentiable rule-based cheminformatics tools (like RDKit) to pre-define rigid substructures and rotatable bonds—failing when encountering complex, ring-rich, or unusual molecules.

Methods

Instead of decomposing a molecule into local fragments or torsion angles, MCF reframes conformer generation as learning a probability distribution over continuous conformer fields $f: \mathcal{G} \to \mathbb{R}^3$, extending Diffusion Probabilistic Fields (DPF) from Euclidean spaces to graphs:

  • Graph Laplacian Positional Encodings: To represent the intrinsic topology of varying molecular graphs $\mathcal{G}$, MCF computes the $k$ leading eigenvectors of the normalized graph Laplacian $\Delta_{\mathcal{G}}$ for each molecule. Each atom vertex $v$ receives an intrinsic coordinate $\phi(v) \in \mathbb{R}^k$.
  • Diffusion Over Functions: Atoms are split into context and query sets. Gaussian noise is injected into the 3D atomic coordinates over discrete diffusion timesteps $t \in [1, T]$, while their intrinsic graph coordinates $\phi(v)$ remain fixed.
  • Score Field Network ($\epsilon_\theta$): A PerceiverIO architecture serves as the score denoiser. Context tokens attend to a latent bottleneck array, which in turn conditions the denoising of arbitrary query tokens.
  • No Explicit Equivariance or Rigid Fragments: MCF learns conformers directly in Cartesian coordinates from data without enforcing hardcoded $SE(3)$ equivariance, bypassing rule-based fragment generators.

Datasets

The authors evaluated MCF across standard benchmarks:

  • GEOM-QM9: $\sim$130k small molecules (3–29 atoms).
  • GEOM-DRUGS: $\sim$304k larger, flexible drug-like molecules (average 44 atoms).
  • GEOM-XL: An out-of-distribution evaluation set containing 102 molecules with $>100$ atoms from GEOM-MoleculeNet to test generalization capability.

Key Insights and Ideas

  • General Function Modeling Overcomes Handcrafted Biases: By avoiding rule-based torsion angle solvers, MCF avoids failure modes on complex topologies. On GEOM-XL, previous state-of-the-art tools failed to generate conformers for 25 out of 102 molecules due to template matching failures; MCF processed all of them seamlessly.
  • Relaxing Explicit Equivariance Works: Deep models can learn global geometric symmetries directly from data when provided with expressive intrinsic coordinates (Laplacian eigenvectors), matching or outperforming computationally rigid equivariant networks.
  • Continuous Interpolation Beyond Atoms: Because the model represents a field rather than a static atom list, it can evaluate points along chemical bonds via linear interpolation of endpoint eigenvectors, yielding physically intuitive continuous electron cloud representations without explicit training on bond paths.

Key Results

  • GEOM-QM9: MCF achieves a 96.57% recall coverage (compared to 92.80% for Torsional Diffusion) and lowers Average Minimum RMSD (AMR) from 0.178 Å to 0.107 Å.
  • GEOM-DRUGS: On large, flexible drug molecules, MCF improves recall coverage from 72.70% to 81.62% and precision coverage from 55.20% to 61.63%, reducing both recall and precision AMR by over 10–20%.
  • GEOM-XL (Generalization): On the subset of 77 molecules where both models succeed, MCF outperforms Torsional Diffusion in precision AMR (1.77 Å vs. 1.93 Å) and recall AMR (2.62 Å vs. 2.84 Å).
  • Fast Sampling: While standard DDPM requires 1000 steps, DDIM sampling achieves comparable quality with only 20 to 50 steps.

Limitations and Future Work

  • High-Energy Outliers: While MCF excels at finding lowest-energy ground states ($E_{\min}$ error of $0.02 \text{ kcal/mol}$), its total ensemble energy error ($E$) is higher than domain-constrained baselines, indicating it occasionally produces high-energy unphysical conformations.
  • Data Dependency: Without explicit physical symmetries, MCF may struggle in extreme low-data regimes or in fine-grained molecular dynamics trajectories.
  • Future Directions: Incorporating flow matching to directly sample from Boltzmann distributions, expanding to protein-conditioned generation (molecular docking), and scaling the functional field paradigm to macromolecular structures like proteins.

Summary updated

Abstract

In this paper we tackle the problem of generating conformers of a molecule in 3D space given its molecular graph. We parameterize these conformers as continuous functions that map elements from the molecular graph to points in 3D space. We then formulate the problem of learning to generate conformers as learning a distribution over these functions using a diffusion generative model, called Molecular Conformer Fields (MCF). Our approach is simple and scalable, and obtains results that are comparable or better than the previous state-of-the-art while making no assumptions about the explicit structure of molecules (\eg modeling torsional angles). MCF represents an advance in extending diffusion models to handle complex scientific problems in a conceptually simple, scalable and effective manner.
BibTeX
@inproceedings{wang2024generating,
  author = {Yuyang Wang and Ahmed A. A. Elhag and Navdeep Jaitly and Joshua M. Susskind and Miguel Ángel Bautista},
  title = {{Generating Molecular Conformer Fields}},
  booktitle = {International Conference on Machine Learning},
  year = {2024},
  url = {https://openreview.net/forum?id=I44Em5D5xy},
}
Figure