Message Passing Neural Network (MPNN)

The message function $M_t$ computes the information sent from a source node to a target node along an edge. It takes the hidden states of both nodes and the edge feature as input.

• Edge-Conditioned: Messages depend on bond type (single, double, aromatic).
• Multiplicative Interactions: Often implemented as a linear layer applied to the concatenated $[h_v, h_w, e_{vw}]$ vector.
• Molecular Context: Encodes local chemical environment, distinguishing a carbon-nitrogen single bond from a carbon-carbon double bond.

The update function $U_t$ computes a new hidden state for each node by combining its current state with the aggregated incoming messages.

• Gated Recurrent Units (GRUs): A common choice, treating the node state as a memory to be updated.
• Residual Connections: Often added to prevent over-smoothing in deep networks.
• Set2Vec: The update must be permutation-invariant to the order of incoming messages, ensuring the model respects graph isomorphism.

The readout function $R$ aggregates all final node hidden states into a single fixed-size graph-level embedding for property prediction.

• Permutation Invariance: Must produce the same vector regardless of node ordering.
• Set2Vec Aggregators: Common choices include sum, mean, and max pooling.
• Set2Vec: An advanced, learned, attention-based readout that weighs node importance.
• Application: Predicts molecular properties like logP, solubility, or binding affinity from the graph embedding.

A critical distinction in geometric deep learning for molecular systems.

• Invariance: The output does not change when the input is rotated. Essential for predicting scalar properties like binding energy.
• Equivariance: The output transforms identically to the input. Essential for predicting vector properties like atomic forces.
• MPNNs are naturally invariant to node permutation but require specialized message functions (e.g., using interatomic distances) to achieve 3D rotational invariance.

A fundamental limitation where node representations become indistinguishable as the number of message-passing layers increases.

• Cause: Repeated local averaging acts as a low-pass filter, collapsing node features to a constant value.
• Impact: Limits the receptive field, preventing the model from capturing long-range dependencies.
• Mitigations: Skip connections, gated update functions, and normalization layers help preserve local information. Graph Attention Networks (GATs) dynamically weight neighbors to combat this.

The parent architecture from which MPNNs are derived. GNNs operate directly on graph-structured data, learning representations by recursively aggregating information from neighboring nodes. Unlike standard neural networks that assume grid-like or sequential input, GNNs respect the permutation invariance of graph data. The MPNN framework formalizes the commonality between various GNN variants, including spectral and spatial approaches, into a unified message-passing paradigm.

An MPNN variant that introduces a self-attention mechanism to compute dynamic, learnable weighting coefficients for each neighboring node's message during aggregation. Rather than treating all neighbors equally or using fixed weights, GATs allow the model to implicitly specify which nodes are most relevant to the target node's updated state. This is critical in drug-target interaction prediction where certain atoms in a binding pocket dominate the non-covalent interaction energy.

A specialized architecture ensuring that predictions transform predictably under 3D rotations and translations of the input molecular structure. Standard MPNNs operating on distance features are invariant but not equivariant to orientation. Equivariant networks, such as tensor field networks and SE(3)-Transformers, preserve directional information through the network layers, enabling more accurate prediction of vectorial properties like atomic forces in molecular dynamics.

The umbrella term for neural architectures designed to respect the symmetries and invariances of non-Euclidean data domains. This includes:- Graphs: Permutation invariance- Grids: Translation equivariance (CNNs)- Manifolds/3D structures: Rotation and translation equivarianceMPNNs are a specific instance of geometric deep learning applied to graph domains. The framework provides the mathematical blueprint for building networks that exploit the known geometry of the input space.

A heterogeneous bipartite graph where drug nodes and protein target nodes are connected by edges representing known or predicted binding interactions. MPNNs process these networks to perform link prediction—inferring missing edges that represent novel therapeutic interactions. The message-passing paradigm naturally captures polypharmacology effects, where a single drug's representation is influenced by all the targets it is known to bind, and vice versa.

A machine learning paradigm that simultaneously models the ligand space and the target space using explicit descriptors for both. MPNNs extend PCM by learning continuous vector representations for both drugs and proteins directly from their structural graphs. This allows the model to generalize to previously unseen targets or ligands—a capability known as zero-shot interaction prediction—by embedding them into the same learned latent space where interaction scores are computed.