3D Geometric Representations for Molecular Neural Networks

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 21,120 papers1,413 conceptsfrozen 2026-10-04

For decades, computational chemistry represented molecules as simplified abstractions: 1D strings (SMILES) or 2D topological graphs. While these formats are efficient for database storage, they ignore a fundamental physical reality: molecules are dynamic, three-dimensional objects. Their behavior—how they fold, how they vibrate, and how they dock into protein pockets—is dictated by their spatial geometry.

A growing body of research is moving beyond "topology-only" models to integrate explicit 3D geometric information. This shift allows neural networks to capture the physical constraints and energy landscapes that govern molecular interactions, leading to more accurate property prediction and more viable drug design.

The 2D Bottleneck and the Importance of Conformation

A 2D graph tells you which atoms are connected, but it doesn't tell you how the molecule sits in space. A single 2D graph can correspond to many different 3D "conformations" (shapes), and the specific conformation a molecule adopts determines its chemical activity. As Sehan Lee, Jaechang Lim, and Woo Youn Kim demonstrate with C3Net, incorporating multiple 3D spatial conformations significantly improves the prediction of physicochemical properties compared to models relying on 2D structures alone.

The challenge is that high-quality 3D data is computationally expensive to generate using traditional quantum mechanics. To solve this, Zhao Xu and colleagues developed Molecule3D, a benchmark designed to train models to predict 3D ground-state geometries directly from 2D graphs. They found that even approximate, machine-learned 3D geometries provide a superior foundation for downstream tasks than 2D connectivity alone, as they capture the underlying physical relationships inherent in spatial proximity.

Encoding Physical Symmetries with SE(3) Awareness

When modeling molecules in 3D space, a model must respect the laws of physics—specifically, it must be invariant or equivariant to rotations and translations (the SE(3) group). If you rotate a molecule, its internal energy doesn't change (invariance), but its atomic coordinate vectors should rotate accordingly (equivariance).

Researchers have developed specialized architectures to handle these symmetries:

  • 3D Transformers: Fang Wu and team introduced Molformer, which treats molecules as 3D heterogeneous graphs. It uses a 3D-aware self-attention mechanism to process the relationships between individual atoms and larger functional motifs while maintaining geometric awareness.
  • Geometric Pretraining: To leverage large-scale unlabeled data, Hongyu Guo and Shengchao Liu proposed GeoSSL-DDM. This framework pretrains models by denoising atomic distances, effectively teaching the AI to understand the "potential energy surface" of a molecule. Similarly, Rui Jiao and colleagues developed 3D-EMGP, which uses an equivariant energy-based backbone to ensure the model respects physical symmetries during pretraining.

Generative Design in Continuous 3D Space

The ultimate goal of this research direction is structure-based drug design: creating a molecule that fits perfectly into a specific biological target. Traditional generative models often produce 1D strings or 2D graphs that, while chemically valid, may be physically impossible to dock into a protein's binding pocket.

By moving to 3D, models can "feel out" the empty space within a target. Shitong Luo introduced a 3D generative model that uses auto-regressive sampling in continuous space to place atoms precisely within the irregular geometry of a protein pocket. This approach avoids the "dimensionality gap" of 2D models, ensuring that generated candidates respect the complex spatial constraints of the biological environment.

Furthering this, Tony Shen and others developed 3DSynthFlow, which co-designs a molecule's 3D binding pose alongside its synthetic pathway. This ensures that the resulting drug is not only geometrically optimized for binding but is also practically manufacturable in a lab.

Why 3D Works: Intuition

The shift to 3D works because it aligns the model's internal representation with the physical forces that actually drive chemistry. A 2D graph is a topological map; it tells you who is neighbors with whom. A 3D representation is a physical model; it tells you how much room a molecule takes up and where its electrons are likely to be. By training models to "see" in 3D, we allow them to learn the implicit rules of steric hindrance, van der Waals forces, and electrostatic interactions that 2D models can only approximate through indirect patterns.

Papers in this version

Showing 60 of 1,120 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 632 papers
  2. Version 1 · 2026-10-04 · 1120 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_geometric_representation_gap_61cf3a6e,
  title = {geometric representation gap},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/geometric-representation-gap/rdr_61cf3a6e707e41beb6284c610afea3e5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_3d_geometric_representations_for_molecul_61cf3a6e_v1,
  title = {3D Geometric Representations for Molecular Neural Networks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/3d-geometric-representations-for-molecular-neural-networks/rdr_61cf3a6e707e41beb6284c610afea3e5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}