📋 KEY INSIGHTS
- Geometric deep learning (GDL) is the unifying framework that extends deep learning to data with non-Euclidean structure โ graphs, point clouds, meshes, manifolds, and 3D geometric objects โ by building neural network architectures that respect the symmetries of the underlying data domain. The core principle: a learning algorithm should be equivariant to the symmetries of its input domain. For images, translational equivariance gives CNNs. For sets, permutation equivariance gives DeepSets and self-attention. For molecules and 3D point clouds, E(3)-equivariance (invariance to rotations, reflections, and translations) is essential โ a molecule does not change its properties when rotated in 3D space.
- Group theory formalises symmetries for GDL: a group G acts on the input space X, and an equivariant function f satisfies f(gยทx) = gยทf(x) for all g in G and x in X (the output transforms the same way as the input under the group action). An invariant function satisfies f(gยทx) = f(x) โ the output is unchanged by the symmetry. For molecular property prediction, we want invariant models (properties do not change under rotation); for force prediction, we want equivariant models (forces are vectors that rotate with the molecule). Key groups: E(3) = rotations + reflections + translations of R^3; SE(3) = rotations + translations only; O(3) = rotations + reflections.
- E(3)-equivariant GNNs build on standard message passing by using geometric quantities (interatomic distances, angles, dihedral angles) as edge features and ensuring that node features transform correctly under rotation and reflection. SchNet uses only radial basis functions of interatomic distances (invariant features), DimeNet adds angular information (also invariant), and SEGNN/MACE use spherical harmonic representations of relative position vectors (equivariant features) that enable the network to reason about 3D geometry directly. MACE (Message Passing Neural Network with Equivariant Many-Body Interactions) achieves state-of-the-art accuracy on the MD17 molecular dynamics benchmark with dramatically fewer parameters than transformer-based alternatives.
- PointNet (Qi et al., 2017) is the foundational architecture for learning on 3D point clouds โ unordered sets of (x, y, z) coordinates representing surfaces sampled from 3D objects. PointNet achieves permutation invariance by applying a shared MLP to each point independently, then using a symmetric aggregation function (global max pooling) to produce a single global feature. PointNet++ extends this with hierarchical local feature learning: farthest point sampling selects representative points, and ball query groups nearby points into local neighbourhoods for local PointNet processing โ analogous to a 3D CNN but without voxelisation.
- The Geometric Algebra Transformer (GATr, Brehmer et al., 2023) and SE(3)-Transformer (Fuchs et al., 2020) build Transformer architectures with geometric equivariance by operating on geometric algebra multivectors (which transform correctly under E(3)) instead of standard vector embeddings. This enables attention-based architectures to process 3D point clouds and molecular structures while maintaining equivariance, combining the expressiveness of multi-head attention with the geometric correctness of equivariant networks. The connection to our Transformers guide is direct: equivariant Transformers extend standard Transformers with geometric structure, using the same attention mechanism but with equivariant feature representations.
- Practical GDL tooling: PyTorch Geometric (PyG) provides the most comprehensive GDL framework โ including equivariant convolutions, E(3)GNN implementations, point cloud operations, and mesh processing โ via a unified
MessagePassingbase class that handles heterogeneous graph types. DGL (Deep Graph Library) is an alternative with strong support for distributed training on large graphs. The e3nn library provides PyTorch primitives for E(3)-equivariant neural networks: spherical harmonic tensor products, Wigner D-matrices for feature rotation, and equivariant linear layers โ the building blocks from which SEGNN, NequIP, and MACE are assembled.
Standard deep learning architectures โ CNNs for images, Transformers for sequences โ achieve their power by exploiting the symmetries of their data domains: CNNs are translationally equivariant because images are translation-invariant (a cat is a cat regardless of where in the image it appears), and this inductive bias dramatically reduces the effective number of parameters needed to learn from image data. When we move to 3D geometric data โ molecular structures, protein conformations, robot poses, 3D scene point clouds โ the relevant symmetries are rotational, reflective, and translational (the E(3) group), and standard architectures fail to encode these symmetries. Geometric deep learning provides architectures that are equivariant to these geometric symmetries by design, achieving better sample efficiency, generalisation, and physical interpretability than non-equivariant baselines. This guide covers the mathematical foundations of geometric deep learning, E(3)-equivariant GNNs for molecular modelling, PointNet and PointNet++ for point cloud processing, equivariant Transformers, and practical tooling with PyTorch Geometric and e3nn.
Equivariance, Group Theory and Molecular GNNs
The mathematical foundation of geometric deep learning is the equivariance principle from group theory. For a network operating on 3D molecular structures, equivariance to SE(3) (special Euclidean group: rotations and translations of R^3) means: if we rotate a molecule by a rotation matrix R and translate it by a vector t, the network output transforms predictably โ scalar properties (energy, binding affinity) are invariant (unchanged), vector properties (atomic forces, dipole moment) transform as Rยทoutput, and tensor properties (stress tensor) transform as RยทoutputยทR^T. Building equivariance into the architecture rather than enforcing it via data augmentation is vastly more parameter-efficient: data augmentation requires training on many rotated copies of each molecule, while an equivariant network handles all rotations with a single forward pass. SchNet, one of the first successful equivariant molecular GNNs, achieves invariance simply by using only interatomic distances as edge features โ Euclidean distances are rotation-invariant by construction, so any function of distances is automatically invariant. DimeNet extends this with bond angles (also invariant), enabling better discrimination of molecular geometry. MACE (Batatia et al., 2022) achieves a qualitative leap by using equivariant spherical harmonic features: each atom carries a feature vector in an irreducible representation of O(3) (a direct sum of spherical harmonics of degrees l=0,1,2,…), and message passing uses equivariant tensor products (Clebsch-Gordan products) to combine features from neighbouring atoms in a geometrically consistent way. The connection to our drug discovery guide is direct: MACE and NequIP are the state-of-the-art force field models for molecular dynamics simulation, enabling molecular simulations with near-DFT accuracy at a fraction of the computational cost.
| Architecture | Symmetry | Features | Benchmark | Primary Use |
|---|---|---|---|---|
| SchNet | E(3)-invariant | Distances only | MD17 (baseline) | Molecular energy/forces |
| DimeNet++ | E(3)-invariant | Distances + angles | OC20, MD17 | Catalyst property prediction |
| MACE | E(3)-equivariant | Spherical harmonics | MD17 SOTA | Force fields; molecular dynamics |
| PointNet | SE(3)-invariant (via global max pool) | Raw XYZ coordinates | ModelNet40 | 3D object classification |
| PointNet++ | SE(3)-invariant | Hierarchical local features | ShapeNet parts | 3D segmentation; LiDAR |
| SE(3)-Transformer | SE(3)-equivariant | Spherical harmonics + attention | QM9 | Molecular property; protein structure |
PointNet, Point Cloud Processing and Equivariant Transformers
Point clouds โ unordered sets of 3D points sampled from surfaces โ are the native output of LiDAR sensors (self-driving cars, robotics) and 3D scanners. The core challenge: unlike images (regular pixel grids) or graphs (defined connectivity), point clouds have no inherent ordering and variable density across the surface. PointNet addresses this by exploiting two properties: (1) any symmetric function of the point set is permutation-invariant, and (2) any continuous symmetric function can be approximated by a composition of a point-wise function (applied independently to each point) and a symmetric aggregation. The PointNet architecture: apply a shared MLP (implemented as 1D convolution) to each point independently, producing per-point features; global max-pool across all points to produce a permutation-invariant global descriptor; then use the global descriptor for classification or segmentation. The connection to our neural network architectures guide is direct: PointNet is architecturally a DeepSet โ a permutation-invariant set function implemented as point-wise MLP + symmetric aggregation, which is the set-function analogue of the image CNN. PointNet++ adds spatial hierarchy: farthest point sampling selects N representative points from the full cloud, ball query groups the K nearest points around each representative into a local patch, and PointNet is applied to each local patch โ analogous to a 3D convolutional layer with learnable receptive field. The SE(3)-Transformer (Fuchs et al., 2020) replaces standard attention (which operates on scalar dot products between vector embeddings) with equivariant attention: attention weights are computed from invariant pairwise features (distances, angles), and value aggregation uses equivariant tensor products that preserve the geometric transformation law. This enables the Transformer architecture to process 3D point clouds and molecular structures while maintaining SE(3)-equivariance, combining the multi-head attention expressiveness of our Transformers guide with the geometric correctness of equivariant GNNs. Applications include protein structure refinement (complementing AlphaFold2 from our drug discovery guide), rigid body pose estimation, and scene flow estimation from LiDAR point clouds for autonomous driving.
✦ SUMMARIZE THIS ARTICLE WITH AI
ChatGPTGraph Neural Networks (the message passing framework underlying molecular GNNs) are in our GNN guide. Graph Attention Networks (the equivariant attention variant) are in our Graph Attention Networks guide. ML for drug discovery (MACE and NequIP for force fields) is in our Drug Discovery guide. Transformers (the architecture extended by SE(3)-Transformers) are in our Transformers guide. Neural ODEs (continuous geometric dynamics) are in our Neural ODEs guide. Topological data analysis (complementary geometric ML method) is in our TDA guide.
