Graph Neural Network (GNN)

The foundational mechanism of most GNNs, where nodes iteratively update their hidden states by aggregating feature information from their local neighborhood.

• Message Function: Computes a message vector from a neighbor node's current state and the edge features connecting them.
• Aggregation Function: A permutation-invariant operation (e.g., SUM, MEAN, MAX) that combines all incoming messages.
• Update Function: Combines the aggregated message with the node's previous state to produce a new representation.

This process is repeated for K layers, allowing each node to integrate information from its K-hop neighborhood.

GATs introduce a self-attention mechanism to the message-passing framework, allowing the model to learn the relative importance of each neighbor during aggregation.

• Dynamic Weighting: Instead of treating all neighbors equally, GAT computes an attention coefficient for each edge.
• Mechanism: A shared, learnable linear transformation is applied to node features, followed by a single-layer feedforward network that outputs attention scores.
• Multi-Head Attention: Multiple independent attention mechanisms are often run in parallel and concatenated to stabilize learning.

This is particularly useful in drug-target interaction networks where different binding interactions have varying degrees of pharmacological relevance.

For molecular applications, standard GNNs are insufficient because they treat molecules as 2D graphs. Equivariant GNNs respect the 3D symmetries of physical systems.

• SE(3) Equivariance: The model's predictions must rotate and translate consistently with the input atomic coordinates. If you rotate a protein-ligand complex, the predicted binding affinity must remain identical.
• Tensor Field Networks: Use spherical harmonics and Clebsch-Gordan tensor products to build features that transform predictably under rotation.
• Key Benefit: Eliminates the need for data augmentation by rotation, as the architecture mathematically guarantees orientation-independent predictions for molecular dynamics and docking scores.

GNNs are flexible architectures that can be configured for three distinct prediction tasks, all relevant to drug-target interaction prediction.

• Node Classification: Predict a property for each node, such as identifying a protein's binding site residues or classifying atoms by element type.
• Edge Prediction (Link Prediction): Predict the existence or strength of a connection between two nodes. This is the core task for predicting novel drug-target interactions in a bipartite graph.
• Graph Classification/Regression: Predict a global property for the entire graph, such as a molecule's toxicity or a protein-ligand complex's binding affinity.

A readout function (global pooling) aggregates all node states to produce a graph-level embedding.

Unlike transductive methods (e.g., Node2Vec) that require retraining for new nodes, GNNs are inherently inductive.

• Shared Parameters: The learned message, aggregation, and update functions are parameterized by the graph's topology, not by specific node IDs.
• Generalization: A GNN trained on a set of known protein-ligand complexes can immediately predict interactions for a completely novel drug candidate or a newly crystallized target protein without retraining.

This is critical for virtual screening, where the goal is to rank millions of molecules never seen during training against a specific target.

GINs are designed to achieve maximum discriminative power, theoretically equivalent to the Weisfeiler-Lehman (WL) graph isomorphism test.

• Limitation of MEAN/MAX: Standard aggregators like MEAN or MAX pooling can fail to distinguish between simple graph structures that are non-isomorphic.
• SUM Aggregation: GIN uses SUM aggregation combined with a multi-layer perceptron (MLP) update function to learn injective neighborhood representations.
• Application: In molecular fingerprinting, GINs can distinguish between different chemical structures that simpler GNNs would embed identically, leading to more accurate QSAR models.

A heterogeneous bipartite graph where nodes represent drugs and protein targets, and edges encode known or predicted interactions. GNNs operating on these networks perform link prediction to identify novel drug-target pairs. By incorporating drug-drug similarity, target-target similarity, and known interaction edges into a unified graph, these models enable systems-level polypharmacology analysis and drug repurposing hypothesis generation.

A fixed-length bit-string or vector encoding the presence or absence of specific substructural features within a molecule. Traditional fingerprints like ECFP4 use circular hashing algorithms. GNNs can learn differentiable, task-specific fingerprints by backpropagating through the message-passing layers, producing continuous vector representations that often outperform hand-crafted fingerprints on property prediction benchmarks.