GNN for Science: Foundations, Applications, and Recent Advances in Equivariant Graph Neural Networks

Machine learning model interpretability has become a major focus, and graph machine learning, as a crucial component of AI, demands deeper study of its explainability. Integrating scientific knowledge into machine learning, as exemplified by GNN for Science, offers new insights into model interpretability.

The talk introduces AI for Science as a turning point, highlighting how large‑scale pre‑training models (e.g., GPT‑3, ViT) and breakthroughs like AlphaFold2 have expanded AI from classic tasks to scientific domains such as biology and physics.

Graphs naturally arise in scientific data: molecules, proteins, and celestial bodies can all be modeled as graphs where nodes represent entities and edges represent interactions. Modeling these data as graphs enables the use of graph neural networks (GNNs) for representation and analysis.

GNNs have a rich history, beginning with early work in 1997 and gaining rapid attention from 2014 onward. Two main development streams—RNN‑based and graph‑signal‑processing‑based—converge in the message‑passing neural network (MPNN) framework, which is now the most widely used GNN model.

MPNNs are illustrated with a four‑node example, showing the aggregation (AGG) of neighbor features and the update (U) of the target node. While MPNNs are simple and versatile, many other GNN variants exist, such as the geometry‑aware models.

Applications in physics include modeling chaotic systems like double pendulums and simulating fluid dynamics by treating particles as graph nodes and their interactions as edges. GNNs can predict system states at future time steps, as demonstrated by recent works (e.g., Sanchez‑Gonzalez et al., 2020).

In biomedicine, GNNs are applied to drug discovery, target identification, and interaction prediction. Large‑scale self‑supervised pre‑training on unlabeled molecular graphs (e.g., the GROVER model with 100 M parameters) significantly improves property prediction.

Recent research focuses on Euclidean equivariant GNNs, which respect physical symmetries such as rotation invariance. By embedding symmetry constraints into the model (e.g., using scalar, vector, and tensor representations), these networks achieve equivariance and improve performance on scientific tasks.

The authors’ own method, GMN, extends MPNN by incorporating full‑directional forces through a matrix of geometric information, enabling accurate simulation of physical systems and molecular dynamics.

Additional work includes advances in protein dynamics simulation, antibody generation, and optimization, where incorporating physical laws leads to reduced prediction errors.

In summary, GNN for Science differs from traditional GNNs by: (1) handling multi‑dimensional data (1D/2D/3D/4D); (2) integrating domain knowledge from physics, chemistry, and biology; (3) addressing a broader range of scientific applications; and (4) fostering interdisciplinary research.

The presentation concludes with thanks to the audience.