Coordinate Denoising for Geometric Representation Learning

57 concepts41 papers1 proposal

In fields like drug discovery and materials science, predicting how a molecule or protein behaves requires a deep understanding of its 3D geometry. However, labeled data—such as experimentally determined binding affinities or quantum mechanical properties—is notoriously scarce and expensive to compute.

To overcome this data bottleneck, researchers have turned to self-supervised pre-training. While text models learn by masking words, and vision models learn by masking pixels, geometric models learn by coordinate denoising. The core mechanism is simple: take a stable 3D molecular structure, corrupt the spatial coordinates of its atoms with random noise, and train a neural network to reconstruct the original, uncorrupted positions.

As demonstrated by Gengmo Zhou et al. in Uni-Mol: A Universal 3D Molecular Representation Learning Framework, this simple task acts as an incredibly powerful proxy for learning chemical physics. But why does removing artificial noise teach a model about real-world chemistry?

The Mathematical Equivalence to Physical Forces

To understand why coordinate denoising works, we must look at the relationship between a molecule's geometry and its energy. In nature, molecules are not static; they settle into stable, low-energy configurations known as equilibrium conformations. If you manually push an atom out of its equilibrium position, the chemical bonds act like tiny springs, exerting a physical restorative force to pull the atom back into place. The landscape mapping a molecule's spatial coordinates to its potential energy is called the Potential Energy Surface (PES).

When we train a model to perform coordinate denoising, we are implicitly forcing it to learn the topology of this energy landscape. Sheheryar Zaidi proves in Pre-training via Denoising for Molecular Property Prediction that coordinate denoising with Gaussian noise is mathematically equivalent to denoising score matching.

Score matching is a technique designed to estimate the gradient of a data distribution's log density. Because stable molecular conformations dominate the training data, the data distribution peaks at these low-energy states. The gradient of this distribution points directly toward the nearest stable configuration. In physical terms, this gradient corresponds exactly to the conservative forces acting on the atoms.

By learning to predict the noise vector required to restore an atom to its original position, the model is actually learning to predict the "virtual forces" that govern molecular stability. As Rui Jiao et al. show in Energy-Motivated Equivariant Pretraining for 3D Molecular Graphs, framing denoising as a virtual force prediction task bridges the gap between generative machine learning and classical physics, allowing models to internalize structural dynamics without ever seeing an explicit physical force label. Rui Feng leverages this insight in May the Force be with You: Unified Force-Centric Pre-Training for 3D Molecular Conformations to successfully model both stable equilibrium states and high-energy, off-equilibrium dynamics.

Refining the Noise: Physics-Informed Perturbations

Standard coordinate denoising typically applies isotropic Gaussian noise, jiggling every atom independently in 3D space. While mathematically convenient, this approach ignores basic chemical realities. Shaking atoms independently can easily break rigid chemical invariants, resulting in intermediate corrupted states with physically impossible bond lengths or overlapping atoms.

To force models to respect true chemical constraints, recent work introduces structured, physics-informed noise distributions. For instance, Shikun Feng et al. introduce Fractional Denoising for 3D Molecular Pre-training, which perturbs complex structural features like dihedral angles but only requires the model to resolve the resulting coordinate displacements. This partial denoising strategy acts as a targeted data augmentation that preserves a valid physical force field interpretation.

Taking this concept a step further, Yuyan Ni et al. propose Sliced Denoising: A Physics-Informed Molecular Pre-Training Method. Instead of applying arbitrary spatial noise, the authors explicitly perturb the internal coordinates of the molecule—its bond lengths, bond angles, and torsional angles. By training the model to denoise these specific geometric descriptors, the network develops a much finer appreciation for classical mechanical principles, resulting in massive accuracy gains when predicting molecular forces.

Scaling to Macromolecules

When moving from small molecules to massive macromolecular structures like proteins, atom-level coordinate denoising begins to struggle. Jiggling individual atoms in a protein backbone provides a weak learning signal for capturing the large-scale, global topology essential for protein folding and binding.

To scale coordinate denoising to the protein universe, Rui Jiao et al. introduce a block-level denoising task in Equivariant Pretrained Transformer for Unified Geometric Learning on Multi-Domain 3D Molecules. Instead of perturbing individual atoms, they apply random rotations and translations to entire structural blocks as rigid bodies. The model must then predict the restorative torques and forces needed to realign the blocks. This preserves local chemical validity (like individual amino acid structures) while forcing the network to capture complex, long-range spatial relationships.

Standard Denoising:   [Atom 1] ➔ ✦ (Isotropic Jiggling)
                      [Atom 2] ➔ ✦

Block Denoising:      [ Amino Acid Residue ] ➔ ⟳ ➔ (Rigid Rotation/Translation)
                      [ Amino Acid Residue ] ➔ ⟲

Coordinate denoising also proves highly effective at correcting systematic distribution shifts in synthetic structural data. For example, structural prediction tools like AlphaFold generate highly accurate protein structures, but these outputs are often "too perfect," lacking the natural physical variability and geometric nuances found in real experimental data. Cheng Tan et al. address this in AlphaFold Database Debiasing for Robust Inverse Folding by training a denoising autoencoder to map corrupted synthetic backbones back onto the realistic physical manifold of experimental structures, significantly improving the robustness of downstream protein design tools.

Cross-Modal Knowledge Transfer

A major practical hurdle for 3D geometric models is that 3D coordinates are rarely available during real-world inference. If a chemist wants to screen millions of candidate molecules, generating accurate 3D conformations for all of them is computationally prohibitive; they usually only have access to simple 2D molecular graphs.

Coordinate denoising elegantly solves this problem through cross-modal distillation. As explored by Sungjun Cho et al. in 3D Denoisers Are Good 2D Teachers: Molecular Pretraining via Denoising and Cross-Modal Distillation, we can pre-train a sophisticated 3D "teacher" model using coordinate denoising to master molecular physics. Once the teacher understands the energy landscape, its spatial representations are distilled into a 2D "student" graph neural network.

Through this distillation process, the 2D model learns to implicitly "imagine" the underlying 3D geometric constraints and physical laws of a molecule. At inference time, the model achieves the high performance associated with 3D structural data while only requiring a simple 2D graph as input, bypassing the need for expensive 3D coordinate generation entirely.

Key Foundational Work

A Systematic Survey of Molecular Pre-trained Models (2022) — contributed 1 concept elements to this direction.
Crystal Diffusion Variational Autoencoder for Periodic Material Generation (2022) — contributed 1 concept elements to this direction.
Recombinator Networks: Learning Coarse-to-Fine Feature Aggregation (2016) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_coordinate_denoising_4a8a3160,
  title = {coordinate denoising},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/coordinate-denoising/rdr_4a8a316064b3434e9c9fd8b1cd6fa926},
  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_coordinate_denoising_4a8a3160_v1,
  title = {coordinate denoising},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/coordinate-denoising/rdr_4a8a316064b3434e9c9fd8b1cd6fa926/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake