Applications of Graph Neural Networks in Biochemistry and Drug Discovery

Graph Neural Networks (GNNs) have attracted extensive attention due to their rich application scenarios, especially in biochemistry and medical domains.

In molecular contexts, atoms are treated as nodes and chemical bonds as edges, allowing tasks such as molecular property prediction (e.g., classification of MUTAG, toxicity in Tox21, anticancer activity in NCI‑1) by learning graph‑level representations often called “molecular fingerprints”. Early works like Duvenaud et al. and Gilmer et al. introduced convolutional and message‑passing approaches that significantly advanced these predictions.

Chemical reaction prediction is framed as a graph‑to‑graph problem: given a disjoint graph of reactant molecules, a GNN encodes each atom, scores possible atom‑pair interactions, selects top‑scoring pairs, enumerates candidate products, and re‑ranks them with another GNN, as illustrated in the accompanying diagram.

Beyond prediction, GNNs drive molecular generation. Autoregressive models such as GraphRNN generate graphs node‑by‑node based on a predefined ordering. Generative adversarial networks (e.g., MolGAN) map latent vectors to adjacency tensors and node attribute matrices, though they may suffer from mode collapse and lack of chemical validity constraints.

Variational auto‑encoders (VAEs) are employed for molecule generation, with examples like Junction‑Tree VAE (JT‑VAE) that first decompose a molecule into a tree of substructures and then decode the tree and graph jointly. Constrained VAE variants add mask‑based or rule‑based restrictions to enforce valence and connectivity.

Normalizing‑flow models such as Real NVP and GraphNVP provide invertible mappings between latent vectors and graph representations, enabling exact likelihood training and reversible generation of molecular graphs.

Formally, given known molecular graphs and their properties sampled from an unknown distribution p(G) , a graph generative model aims to learn this distribution and sample new valid molecular graphs with desired chemical attributes.

GNNs also process protein‑protein interaction graphs, drug‑drug interaction (DDI) graphs, and drug‑target interaction (DTI) graphs for link prediction, facilitating safer drug recommendation. Models like GAMENet combine a GNN‑encoded DDI graph with a memory network to avoid adverse drug combinations when recommending medications.

The article concludes with a list of representative references covering convolutional, message‑passing, reaction‑prediction, generative, and application‑oriented GNN research.