When modeling the 3D structure of a molecule, the most obvious approach is to track the $(x, y, z)$ coordinates of every atom. However, this Cartesian representation is highly redundant. Atoms in a molecule are not free-floating; they are bound by covalent forces that keep bond lengths and bond angles relatively rigid. The true "flexibility" of a molecule—its ability to shift between different shapes, or conformers—primarily arises from rotations around its single bonds.
A growing body of research suggests that by shifting from Cartesian coordinates to internal coordinates (specifically torsion angles), we can build models that are more efficient, physically accurate, and easier to train.
The Geometry of Reduced Degrees of Freedom
In a molecule with $N$ atoms, a Cartesian representation requires $3N$ variables. However, most of these variables are constrained by local chemistry. As Tremayne et al. noted in early work on crystal structures, modeling molecules as rigid fragments with specific internal degrees of freedom significantly reduces the dimensionality of the search space.
This reduction is dramatic. Jing et al. demonstrate that for typical drug-like molecules, focusing only on rotatable bonds reduces the degrees of freedom by over 90%. By treating bond lengths and angles as fixed constants—often derived from standard cheminformatics tools as described by Zhang et al.—models can focus their entire computational budget on the torsion angles that actually define the molecule's shape.
Guaranteed Chemical Validity
One of the primary failures of Cartesian-based generative models is the production of "unphysical" structures, such as atoms that are too far apart or bonds that are bent at impossible angles. To fix these, researchers often have to use computationally expensive post-processing or force-field minimization.
Internal coordinates solve this by design. Because the model only manipulates angles while keeping bond lengths fixed, the resulting 3D coordinates naturally adhere to chemical constraints. Misra et al. showed that this "perfect chemical validity" allows torsional models to outperform much larger unconstrained models while using 60x fewer parameters. Similarly, Ganea et al. argue that predicting exactly one torsion angle per non-terminal bond avoids the over-parameterization issues of distance-matrix methods, which often produce physically impossible geometries.
Generative Modeling on the Hypertorus
Shifting to internal coordinates changes the mathematical "space" the model lives in. Instead of Euclidean space, torsion angles exist on a hypertorus—a multi-dimensional surface where an angle of $0^\circ$ is the same as $360^\circ$.
Recent breakthroughs have applied modern generative frameworks to this angular space:
- Torsional Diffusion: Jing et al. and Corso et al. developed diffusion processes that add and reverse noise directly on the torsion angles. This allows for faster sampling and better recovery of global energy minima.
- Flow Matching: Lin et al. introduced PPFLOW, which treats peptide design as a fluid movement through the space of molecular "joints" on a torus manifold.
- GFlowNets: Volokhova et al. used Generative Flow Networks to sample conformations proportional to their physical probability (the Boltzmann distribution) by focusing exclusively on rotatable torsion angles.
Scaling to Complex Biopolymers
The internal coordinate approach is particularly powerful for large biological molecules like proteins and RNA, where the number of atoms makes Cartesian modeling prohibitive.
For proteins, Ardèvol et al. highlight that backbone dihedral angles provide a compact, rotationally invariant description of topology. In the realm of de novo design, Anand et al. represent complex RNA backbones as a hierarchy of rigid-body frames and predicted torsion angles. This ensures that even as the model generates entirely new structures, they remain within the physically valid range of shapes found in nature, as emphasized by Lin et al..
The Challenge of Long-Range Correlations
While internal coordinates simplify local geometry, they introduce a new challenge: the "lever arm" effect. A tiny change in a single bond angle at one end of a long protein chain can cause a massive, unphysical displacement at the other end.
Arts et al. address this by noting that "covariance matters." To prevent the structure from flying apart, models must account for long-range spatial correlations. While internal coordinates provide inherent rotation and translation invariance, they require specialized handling to ensure that the cumulative effect of many small angular changes results in a globally coherent structure.
By prioritizing these true degrees of freedom, researchers are moving away from treating molecules as clouds of independent atoms and toward a representation that reflects the constrained, articulated reality of chemical bonds.
Key Foundational Work
Tork: Conformational analysis method for molecules and complexes (2003) — contributed 4 concept elements to this direction.
Determination of molecular crystal structures from X-ray powder diffraction by Monte Carlo methods (1996) — contributed 1 concept elements to this direction.