Graph Neural Network (GNN)

The foundational mechanism of GNNs where nodes iteratively exchange information with their neighbors to build contextual representations. In each layer, a node aggregates feature vectors from its local neighborhood, applies a learnable transformation, and updates its own state. This process is repeated for K layers, allowing information to propagate across K-hop neighborhoods. In bioinformatics, message passing enables a protein's representation to incorporate information about its interacting partners, binding sites, and functional annotations. Common aggregation functions include mean, sum, and max pooling, while attention-based aggregators learn to weigh neighbor importance dynamically.

GNNs are designed to be permutation invariant at the graph level and permutation equivariant at the node level. This means the output for an entire graph is unchanged regardless of the arbitrary ordering of nodes in the input adjacency matrix, while node-level predictions permute consistently with the input ordering. This property is critical for molecular graphs, where atoms have no inherent sequential order. Standard architectures like Graph Isomorphism Networks (GINs) achieve maximal discriminative power under this constraint, while Message Passing Neural Networks (MPNNs) provide a general framework that unifies many GNN variants under a common message-passing paradigm.

Graph convolutions generalize the convolution operation from regular grids to irregular graph domains. The Graph Convolutional Network (GCN) applies a first-order approximation of spectral graph convolutions, using a normalized adjacency matrix with self-loops to aggregate neighbor features. More advanced operators include GraphSAGE, which samples and aggregates features from a node's local neighborhood, and Graph Attention Networks (GATs), which learn implicit edge weights via self-attention. In drug discovery, graph convolutions allow models to learn molecular fingerprints directly from atomic connectivity, bypassing hand-crafted feature engineering.

GNNs support learning at multiple granularities. Node-level tasks predict properties of individual entities, such as protein function or atom charge. Edge-level tasks predict relationships, such as drug-target binding affinity or gene regulatory interactions. Graph-level tasks produce a single prediction for an entire graph, such as molecular toxicity or solubility. Graph-level readout is typically achieved via global pooling operations like sum, mean, or more sophisticated hierarchical pooling methods such as DiffPool and Set2Set, which learn to coarsen the graph structure adaptively during training.

GNNs support both transductive learning, where the model sees the entire graph structure during training and only labels for a subset of nodes, and inductive learning, where the model generalizes to entirely unseen graphs or nodes. Inductive capability is essential for drug discovery applications where models trained on known molecular graphs must predict properties of novel, previously unseen compounds. Architectures like GraphSAGE are explicitly designed for inductive settings by learning aggregation functions rather than relying on fixed node embeddings, enabling zero-shot generalization to new molecular scaffolds.

For 3D molecular structures, standard GNNs are extended to incorporate geometric information such as atomic coordinates, bond angles, and interatomic distances. Equivariant GNNs ensure that predictions transform consistently with rotations and translations of the input structure—a physical requirement for molecular property prediction. Architectures like SE(3)-Transformers and Tensor Field Networks operate on spherical harmonics and irreducible representations to maintain strict rotational equivariance. These models are foundational for protein structure prediction, conformer generation, and quantum chemical property estimation.

The fundamental computational mechanism of GNNs where nodes iteratively exchange feature vectors with their neighbors. In each layer, a node aggregates messages from its local neighborhood, updates its own hidden state, and passes a new message outward. This process allows information to propagate across the graph's topology, enabling a protein's residue to learn from its spatial neighbors or a gene to incorporate regulatory signals from connected transcription factors.

A foundational GNN variant that generalizes the convolution operation to irregular graph domains. A GCN layer performs a spectral or spatial convolution by computing a weighted average of a node's features and its neighbors' features, followed by a non-linear activation. In bioinformatics, GCNs are widely used for molecular property prediction and gene-disease association tasks due to their simplicity and effectiveness at capturing local structural patterns.

An advanced GNN architecture that introduces learnable attention coefficients to weigh the importance of different neighboring nodes during message aggregation. Instead of treating all neighbors equally, a GAT dynamically computes how relevant each connection is to the current node's update. This is critical in multi-omics integration where a gene's interaction with one protein may be far more biologically significant than its interaction with another.

The process of mapping discrete graph nodes into a continuous, low-dimensional vector space that preserves structural and feature-based similarities. After training, nodes with similar local topologies or functions—such as proteins with analogous binding pockets—will have embeddings that cluster together. These learned representations serve as powerful input features for downstream tasks like cell-type classification or drug-target interaction prediction.

A GNN variant designed to achieve maximum discriminative power, theoretically as powerful as the Weisfeiler-Lehman graph isomorphism test. GIN uses a multi-layer perceptron on the sum of a node's features and its neighbors' features, ensuring that distinct graph substructures produce distinct embeddings. This property is vital for distinguishing subtly different molecular scaffolds in de novo drug design.

Two distinct learning paradigms for GNNs. Transductive learning requires the entire graph structure at training time and cannot generalize to unseen nodes. Inductive learning, enabled by frameworks like GraphSAGE, learns a function that generates embeddings for previously unseen nodes or entirely new graphs. Inductive capability is essential for virtual screening where a trained model must predict properties for novel, never-before-seen drug candidates.