Molecular Generation using Internal Coordinate Representations

This direction is no longer present in the current build. Its last version stays readable and citable here.

Version 1 of 138 papers46 conceptsfrozen 2026-10-04

When modeling the 3D structure of a molecule, the most obvious representation—a list of Cartesian coordinates $(x, y, z)$ for every atom—is often the most difficult for a neural network to handle. If you rotate a protein by 45 degrees, its Cartesian coordinates change entirely, yet its physical properties remain identical. To learn from these raw coordinates, a model must either be shown every possible orientation via massive data augmentation or utilize complex equivariant architectures that are mathematically difficult to scale.

A growing body of research suggests a more intuitive path: representing molecules through internal coordinates. Instead of asking where an atom is in absolute space, we describe where it is relative to its neighbors using bond lengths, bond angles, and torsion (dihedral) angles.

The Invariance Problem

The primary motivation for shifting to internal coordinates is achieving SE(3) invariance—ensuring that a molecule’s representation does not change when it is rotated or translated. As noted by Shuangli Li and colleagues in their work on SIGN, relying on Cartesian coordinates forces models to "learn" that the same structure in different orientations is the same entity, which significantly increases computational overhead.

By using coordinate-independent descriptors like interatomic distances or polar angles, models like AEScore achieve rotational invariance by design. This allows the network to focus entirely on the chemical environment rather than the camera angle.

Why Internal Coordinates Work

Internal coordinates, often structured as a Z-matrix, decompose a complex 3D shape into three fundamental geometric features:

  1. Bond Lengths: The distance between two bonded atoms.
  2. Bond Angles: The angle formed by three consecutively bonded atoms.
  3. Torsion Angles: The rotation around a central bond formed by four atoms.

In a landmark paper, Frank Noé and his team introduced Boltzmann Generators, which use these coordinates to sample the equilibrium states of proteins. They observed that while Cartesian coordinates are highly coupled and complex, internal coordinates—especially torsion angles—often follow simpler, near-Gaussian distributions. This makes them an ideal "latent space" for generative models like normalizing flows or diffusion models to operate within.

Enforcing Physical Plausibility

Beyond invariance, internal coordinates provide a powerful inductive bias for physical consistency. In Cartesian space, a small numerical error in a model's output might result in a bond stretching to an impossible length or two atoms overlapping.

Generative models like GenZProt, developed by Soojung Yang and Rafael Gómez-Bombarelli, use internal coordinates to solve the "backmapping" problem—restoring atom-level detail to simplified protein models. Because the model predicts angles and lengths rather than positions, the resulting structures inherently respect the molecule's topology. As Colin A. Grambow demonstrates with RINGER, this representation prevents physical violations without requiring the model to learn these constraints from scratch.

Scaling to Large Systems

This approach is particularly effective for long-chain molecules like proteins. FoldingDiff, a model by Kevin E. Wu and James Y. Zou, treats a protein backbone as a sequence of internal angles. By "diffusing" over these angles, the model can generate diverse, designable protein structures using a standard Transformer architecture, bypassing the need for specialized equivariant layers.

Similarly, in the study of chemical reactions, Jian Sun and Xiang Gu developed TS-DFM, which models the "mountain passes" of chemical transitions in interatomic distance space. This allows the model to focus on the fundamental physics of bond breaking and formation, leading to more robust predictions of transition states.

Summary of Intuition

The shift from Cartesian to internal coordinates represents a move from learning symmetries to embedding symmetries. By choosing a representation that is naturally invariant to how a molecule is positioned in space, researchers are able to:

  • Simplify the learning task: Turning complex 3D distributions into simpler, often independent angular distributions.
  • Guarantee stability: Ensuring that generated molecules don't "break" or overlap unphysically.
  • Improve efficiency: Achieving state-of-the-art results with simpler architectures and less data augmentation.

Papers in this version

History & citationVersion 1 · 2026-10-04

No longer present in the current build · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_molecular_generation_using_internal_coor_6c771d7e,
  title = {Molecular Generation using Internal Coordinate Representations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/molecular-generation-using-internal-coordinate-representations/rdr_6c771d7e532f4b4b8e144756017d9a19},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_molecular_generation_using_internal_coor_6c771d7e_v1,
  title = {Molecular Generation using Internal Coordinate Representations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/molecular-generation-using-internal-coordinate-representations/rdr_6c771d7e532f4b4b8e144756017d9a19/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}