Implicit 3D Geometry for 2D Molecular Graphs

This direction is merged into another direction: geometric representation gap. Its last version stays readable and citable here.

Version 1 of 116 papers20 conceptsfrozen 2026-10-04

In molecular machine learning, researchers face a frustrating trade-off. On one hand, 2D graph representations are computationally "cheap" and fast to process, but they ignore the spatial reality of molecules—the bond angles, torsion, and 3D shapes that actually dictate biological activity. On the other hand, 3D models are highly accurate but require expensive quantum simulations or conformer generation to produce the coordinates needed for inference.

A growing body of research is resolving this tension by teaching 2D models to "imagine" 3D space. By using 3D geometric data exclusively during the pre-training phase, these models learn to embed spatial intuition into 2D representations. At inference time, they operate with the speed of a 2D graph model but the "spatial sense" of a 3D one.

The 2D Blueprint and the 3D Reality

A 2D molecular graph is a blueprint: it tells you which atoms are connected, but not how they sit in space. However, the 3D shape of a molecule is not random; it is constrained by the laws of physics. Xu Wang and colleagues argue in 3D PGT that because 3D structure is a function of the 2D graph, a sufficiently powerful model should be able to infer the former from the latter.

The core insight across this field is that we can use 3D data as a "teacher" during training. Once the 2D model learns the underlying relationship between connectivity and geometry, the teacher can be discarded. The resulting model performs inference using only 2D topological data, effectively bypassing the need for explicit 3D coordinate generation at runtime.

Aligning Latent Spaces via Contrastive Learning

One of the most effective ways to instill this 3D intuition is through cross-modal alignment. In this setup, the model consists of two encoders: one for 2D graphs and one for 3D conformations.

Hannes Stärk and the team behind 3D Infomax pioneered a strategy that maximizes the mutual information between these two representations. By forcing the 2D latent vector to be as similar as possible to the 3D latent vector, the 2D encoder learns to capture implicit geometric features—like the likelihood of a certain bond rotating or the overall "globularness" of the molecule.

Similarly, Austin O. Atsango introduced MolCLaSS, which uses 3D shape similarity scores to guide the learning process. Instead of just looking at a single 3D structure, the model learns to recognize when two different 2D graphs would result in similar 3D shapes. This allows the model to perform 3D-aware searches at speeds up to 100,000 times faster than traditional methods.

Beyond Static Shapes to Conformational Ensembles

Molecules are not rigid statues; they are dynamic, vibrating objects that can adopt many different shapes (conformations). Lin Wang and the creators of GeminiMol took the 3D intuition idea a step further by training models on "conformational space profiles."

Instead of teaching the model to imagine one 3D shape, they teach it to understand the entire range of shapes a molecule can adopt. By training a graph neural network to predict the similarity between 3D ensembles using only 2D inputs, the model gains a zero-shot ability to identify drug targets, as it "understands" the molecule's flexibility and binding potential.

Geometric Constraints as Inductive Biases

While some models learn 3D geometry implicitly, others bake it into the architecture. Jinjia Feng introduced TetraGT, which treats bond angles and torsion as explicit tokens in a Graph Transformer. By pre-training on these tetrahedral geometric constraints, the model solves the "handedness" (chirality) problem—the fact that two molecules can be mirror images of each other but have totally different effects in the body.

This approach is echoed in 3DinforCPI by Ngoc-Quang Nguyen, which uses multi-level feature integration to bridge the gap between topological connections and physical 3D structures. By incorporating these priors, the model becomes significantly better at predicting how "novel" compounds—those never seen in training—will interact with proteins.

Scaling to Virtual Screening

The ultimate goal of this research is to make drug discovery faster and more reliable. Robin Winter and researchers at Pfizer developed SAND (Shape-Aware Neural Descriptor) to solve the problem of screening massive molecular libraries.

By mapping 2D structures directly to latent representations of 3D shape, SAND allows researchers to search for molecules that "fit" into a protein pocket without ever having to run a 3D simulation for the millions of candidates in a library. As Namkyeong Lee demonstrates with 3DMRL, this "spatial intuition" can boost performance across dozens of chemical tasks by up to 25%, all while maintaining the lightning-fast inference speed required for modern high-throughput screening.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_implicit_3d_geometry_for_2d_molecular_gr_28d75cfd,
  title = {Implicit 3D Geometry for 2D Molecular Graphs},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/implicit-3d-geometry-for-2d-molecular-graphs/rdr_28d75cfded714d928f968cdd99c98e03},
  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_implicit_3d_geometry_for_2d_molecular_gr_28d75cfd_v1,
  title = {Implicit 3D Geometry for 2D Molecular Graphs},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/implicit-3d-geometry-for-2d-molecular-graphs/rdr_28d75cfded714d928f968cdd99c98e03/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}