Geometric Deep Learning

Molecular properties are intrinsically invariant to rigid transformations—a molecule's binding affinity does not change if it is rotated or translated in space. Geometric deep learning architectures mathematically guarantee this property through equivariance and invariance constraints.

• Equivariance: The model's internal representations transform predictably when the input is rotated, preserving directional information like atomic forces.
• Invariance: The final prediction remains unchanged under rotation, ensuring consistent binding affinity estimates regardless of molecular orientation.
• Key Insight: Standard CNNs or MLPs require massive data augmentation to approximate these symmetries; geometric architectures bake them into the model's mathematical structure, reducing data hunger by orders of magnitude.

Molecules are naturally represented as graphs where atoms are nodes and bonds are edges. Geometric deep learning operates directly on this representation, avoiding information loss from grid-based or sequence-based encodings.

• Node features: Atomic number, hybridization state, partial charge, and mass.
• Edge features: Bond type, interatomic distance, and whether the edge represents a covalent bond or a non-covalent spatial proximity.
• Global features: Molecular weight, logP, or total charge.
• Advantage: Unlike SMILES strings or molecular fingerprints, graph representations preserve the full topological and geometric connectivity, enabling the model to reason about spatial relationships critical for binding.

The core computational primitive in geometric GNNs is equivariant message passing, where nodes update their state by aggregating information from neighbors while respecting 3D geometry.

• Standard MPNN: Messages depend only on node features and scalar distances, losing directional information.
• Equivariant MPNN: Messages incorporate vector features (e.g., relative atomic positions) that rotate correctly with the molecule. Tensor field networks and SE(3)-Transformers use spherical harmonics and Clebsch-Gordan tensor products to achieve this.
• Outcome: The network can predict vector quantities like atomic forces for molecular dynamics simulation or directional properties like dipole moments, all while maintaining rotational equivariance.

The mathematical backbone of geometric deep learning for molecules is the Special Euclidean group SE(3), which describes all possible 3D rotations and translations. Architectures are designed to be SE(3)-equivariant or SE(3)-invariant.

• SE(3)-Invariant: The output (e.g., binding affinity) is unchanged by any rotation or translation of the input coordinates. Used for scalar property prediction.
• SE(3)-Equivariant: The output transforms in a defined way under rotation. Essential for predicting per-atom forces or generating 3D conformations.
• Practical Impact: An SE(3)-equivariant model trained on one protein-ligand orientation generalizes instantly to any other orientation, a property impossible for standard neural networks.

Tensor Field Networks (TFNs) are a foundational equivariant architecture that builds locally equivariant feature maps from point clouds like molecular geometries.

• Mechanism: TFNs use learnable radial functions and spherical harmonic filters to construct convolution operations that map between irreducible representations of SO(3), the rotation group.
• Hierarchy: Lower layers capture simple geometric motifs (bond angles), while deeper layers compose them into complex features (binding pocket shapes).
• Application: TFNs and their successors (SE(3)-Transformers, NequIP) achieve state-of-the-art accuracy on QM9 and MD17 benchmarks for quantum chemistry property prediction, often reaching chemical accuracy with significantly less training data than invariant models.

The choice between invariant and equivariant architectures depends on the prediction task. Applying the wrong constraint can destroy necessary information or waste computational resources.

• Use Invariance for: Binding affinity prediction, toxicity classification, solubility estimation—any task where the output is a single scalar property of the whole molecule.
• Use Equivariance for: Force field prediction, conformational generation, transition state identification—tasks requiring per-atom vector outputs or iterative coordinate updates.
• Hybrid Approach: Modern architectures like EquiDock use invariant layers for coarse scoring and equivariant layers for fine-grained pose refinement, combining the strengths of both paradigms.