Anisotropic Denoising for 3D Molecular Representation Learning

This direction is split into other directions: anisotropic force field modeling. Its last version stays readable and citable here.

Version 1 of 118 papers44 conceptsfrozen 2026-10-04

In the quest to use AI for drug discovery and materials science, researchers must teach models to understand the 3D "shape" of molecules. A dominant strategy is denoising: taking a perturbed molecular structure and training a model to "fix" it, thereby learning the underlying physical forces.

However, a fundamental flaw has emerged in standard denoising models. Most treat atoms as independent points and apply uniform, spherical (isotropic) noise to their coordinates. In the real world, molecules are not collections of loose beads. They are highly constrained structures where moving an atom a fraction of an angstrom along a chemical bond requires massive energy, while rotating that same atom might require almost none.

Recent research has shifted toward anisotropic denoising, a physics-informed approach that recognizes that the "cost" of movement in a molecule depends entirely on direction.

The Failure of Isotropic Noise

Standard generative models, such as Diffusion Probabilistic Models, typically add Gaussian noise to atomic coordinates. As Jiying Zhang and colleagues argue in SubgDiff, this treats atoms as a "bag of independent points," ignoring the stable functional substructures that define chemistry.

When we apply uniform noise, we ask the model to learn a "spherical" energy landscape. But as Shikun Feng and Yuyan Ni point out in their work on Fractional Denoising (Frad), real molecular energy landscapes are jagged and directional. A model trained on isotropic noise struggles to distinguish between a minor thermal vibration and a physically impossible bond stretching. This mismatch prevents models from achieving "chemical accuracy"—the precision needed to reliably predict molecular properties.

Modeling Directional Stiffness

To fix this, researchers are moving the noise from 3D Cartesian space (x, y, z) into the molecule's internal coordinates.

In the SliDe (Sliced Denoising) framework, Yuyan Ni and collaborators perturb bond lengths, bond angles, and torsion angles independently. Because bond lengths are physically "stiffer" than torsion angles, the noise applied to them is much smaller. This ensures that even a "noisy" molecule remains physically plausible, forcing the model to learn the specific directional stiffness of different chemical environments.

Similarly, Xixian Liu and Rui Jiao introduced AniDS, which learns atom-specific covariance matrices. Instead of a circle of noise, each atom is surrounded by an "ellipsoid" of noise that aligns with its local bonds. By adaptively suppressing noise along the axes of stiff covalent bonds, the model gains an inductive bias that mirrors real-world atomic potentials.

From Euclidean Space to Riemannian Manifolds

If the energy landscape of a molecule is highly anisotropic, perhaps we shouldn't be modeling it in flat Euclidean space at all. Jeheon Woo and Woo Youn Kim proposed the Riemannian Denoising Model (R-DM), which operates on a manifold where distances are defined by potential energy changes rather than spatial displacement.

In this view, the "distance" between two atomic configurations isn't just how far the atoms moved, but how much the energy changed. This naturally handles the sensitivity of molecular structures, where a tiny shift in a bond length is "further" (in energy terms) than a large rotation. This approach has allowed models to finally reach the "holy grail" of chemical accuracy, outperforming traditional Cartesian models in structural precision.

Beyond Point Charges

The need for anisotropy extends even to the most basic force fields. Traditional models often treat atoms as point charges—simple spheres of electricity. However, Roseane R. Silva and Jay W. Ponder demonstrated with the HIPPO potential that modeling atoms as overlapping, anisotropic electron densities is essential. By accounting for the fact that electron clouds repel each other differently depending on their orientation, they achieved state-of-the-art accuracy across water phases.

Why It Works

Anisotropic denoising works because it aligns the machine learning objective with the physical reality of the data. By using structure-aware noise—whether through learnable generators like DenoiseVAE or "Energy Sentinels" like MOES-Pred—we prevent the model from wasting capacity on unphysical configurations.

When the noise respects the molecule's internal constraints, the model doesn't just learn where atoms are; it learns the forces that keep them there. This transition from "spatial jiggling" to "physics-aware restoration" is what allows modern 3D molecular models to generalize from simple small molecules to complex materials and proteins.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_anisotropic_denoising_for_3d_molecular_r_2753d69b,
  title = {Anisotropic Denoising for 3D Molecular Representation Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/anisotropic-denoising-for-3d-molecular-representation-learning/rdr_2753d69b21b64bc4bb3067b01b80a3c0},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_anisotropic_denoising_for_3d_molecular_r_2753d69b_v1,
  title = {Anisotropic Denoising for 3D Molecular Representation Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/anisotropic-denoising-for-3d-molecular-representation-learning/rdr_2753d69b21b64bc4bb3067b01b80a3c0/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}