Graph Neural Networks for Molecular and Materials Representation

1,393 concepts923 papers22 proposals

To predict how a molecule behaves, how a drug binds to a protein, or whether a crystal structure is stable, machine learning models must first "see" the chemical system. Historically, researchers relied on fixed-length, manually engineered fingerprints or sequential text strings like SMILES. While sequential strings can capture basic composition, as noted by Yoon et al. in their work on chemical interactions [DeepCCI: End-to-end Deep Learning for Chemical-Chemical Interaction Prediction], they fundamentally struggle to preserve the spatial and topological relationships critical to physical chemistry.

The core insight of graph-based molecular representation is that chemical systems are naturally represented as graphs. By mapping atoms to nodes and chemical bonds or spatial proximities to edges, neural networks can directly ingest the underlying topology of a compound. As emphasized by Laurianne et al., this representation preserves complex connectivity patterns that are completely lost in fixed-length vectors [Applications of Deep-Learning in Exploiting Large-Scale and Heterogeneous Compound Data in Industrial Pharmaceutical Research].

The Inductive Bias of Message Passing

Why do graph neural networks (GNNs) work so well for chemistry? The answer lies in their alignment with physical reality. In a molecule or crystal, an individual atom's properties—such as its partial charge or hybridization state—are dictated by its immediate local environment.

GNNs exploit this via Message-Passing Neural Networks (MPNNs). In an MPNN, each atomic node updates its state by aggregating feature vectors ("messages") from its connected neighbors. As Käser et al. demonstrate when modeling potential energy surfaces, this iterative propagation allows the network to build highly sophisticated representations of local chemical environments [Machine Learning Models of Vibrating H2CO: Comparing Reproducing Kernels, FCHL, and PhysNet]. Similarly, Ko et al. leverage this capability to extract structural features essential for predicting drug-protein bindings [Bayesian neural network with pretrained protein embedding enhances prediction accuracy of drug-protein interaction].

By explicitly encoding atomic connectivity as a structural prior, graph architectures outperform general-purpose models. For instance, Gonzales et al. show that topology-aware graph encodings consistently beat unordered point-cloud approaches because the edges provide a vital inductive bias for learning physical relationships [MatSciML: A Broad, Multi-Task Benchmark for Solid-State Materials Modeling].

Baking in 3D Geometry and Physical Symmetries

A purely topological graph tells you who is connected to whom, but it does not capture where they are in 3D space. Because molecular properties are highly sensitive to conformation, researchers have developed mechanisms to inject geometric constraints into the graph structure.

Distances and Higher-Order Neighbors

The simplest way to incorporate geometry is to encode interatomic distances directly as edge attributes. Fung et al. demonstrate that this allows models to differentiate between structures that share identical graph connectivity but possess distinct spatial arrangements [Benchmarking graph neural networks for materials chemistry]. To capture more complex geometric constraints like dihedral angles, the single-author work by Simm augments the standard molecular graph with auxiliary edges connecting second- and third-order neighbors, allowing message-passing layers to implicitly reconstruct the global 3D shape [A Generative Model for Molecular Distance Geometry].

Explicit Angle Modeling via Dual Graphs

Alternatively, networks can learn angular contributions implicitly or explicitly. The author Wang shows that by aggregating information across multiple hops, a standard GNN can reconstruct the geometric triangles of an atomic neighborhood to deduce bond angles [Symmetry-adapted graph neural networks for constructing molecular dynamics force fields].

However, treating angles as secondary attributes can limit a model's sensitivity to shape. To elevate geometry to a primary feature, Wu introduces a dual-graph representation: one graph models standard atom-bond connectivity, while a second graph explicitly models the relationships and angles between the bonds themselves [Geometry-enhanced molecular representation learning for property prediction]. This dual-graph paradigm is also successfully utilized by Zhong et al. to capture three-body interactions in complex crystal structures [CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling].

[Atom Graph]               [Bond/Angle Graph]
  Node: Atom                 Node: Bond (Edge from Atom Graph)
  Edge: Chemical Bond        Edge: Angle between Bonds

Equivariance and Symmetries

When a molecule rotates or translates in space, its total potential energy remains unchanged (rotational invariance), but its directional force vectors must rotate exactly with the molecule (rotational covariance). If a neural network does not understand these symmetries, it must waste parameter capacity trying to learn them from data.

Anderson et al. solve this by embedding spherical tensors and Clebsch-Gordan decompositions directly into the message-passing mechanics [Cormorant: Covariant Molecular Neural Networks]. This strict physical invariance explains why specialized MPNNs frequently outperform general-purpose architectures like standard Transformers on quantum chemical tasks, as observed by Khrabrov et al. [$\nabla^2$DFT: A Universal Quantum Chemistry Dataset of Drug-Like Molecules and a Benchmark for Neural Network Potentials].

Scaling from Molecules to Crystals and Interfaces

The power of the graph abstraction is its flexibility across physical scales. Because a graph does not require a fixed grid or a specific dimensionality, the exact same mathematical framework can scale from small drug-like molecules to massive periodic structures and multi-component interfaces.

Periodic Crystal Lattices

In solid-state materials science, crystal structures are modeled as infinite periodic graphs where nodes are atoms and edges span across unit cell boundaries. The single-author papers by Arroyave [Phase Stability Through Machine Learning] and Kelvinius [Graph-based machine learning beyond stable materials and relaxed crystal structures] demonstrate how this abstraction allows GNNs to map local atomic environments to thermodynamic stability and phase diagrams. Similarly, Rosen et al. utilize periodic graphs to capture the long-range structural patterns required to predict electronic band gaps in porous metal–organic frameworks [High-throughput predictions of metal–organic framework electronic properties: theoretical challenges, graph neural networks, and data exploration].

Site-Specific and Continuous Space Queries

While early models focused on predicting a single scalar for an entire structure (like total energy), graph representations now enable site-specific and continuous spatial predictions:

Multi-Component Interfaces

When modeling interactions between different structures—such as a small molecule docking into a massive protein—representing every single atom can become computationally prohibitive. To handle these multi-scale interfaces, models adapt the graph structure:

Ultimately, graph-based molecular representation works because it mirrors the physical reality of chemistry. By treating atoms as nodes and structuring their topological, geometric, and symmetric relationships as edges, graph neural networks provide an expressive, scalable language for molecular and materials design.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 958 papers
  2. Version 1 · 2026-10-04 · 843 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_graph_based_molecular_representation_e04b10c4,
  title = {graph-based molecular representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/graph-based-molecular-representation/rdr_e04b10c4b53340eb8abf65a82fee23ef},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_graph_based_molecular_representation_e04b10c4_v2,
  title = {graph-based molecular representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/graph-based-molecular-representation/rdr_e04b10c4b53340eb8abf65a82fee23ef/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake