Graph Neural Networks for Binding Affinity Prediction

Binding affinity—the strength of a ligand’s interaction with a target biomolecule—is a central metric in early drug discovery because it helps rank candidate “hits” and guide designs that bind tightly to the intended target while avoiding off-target binding that can trigger side effects. Experimentally measuring binding affinity (often via the equilibrium inhibition constant, Ki) is accurate but costly in time, money, and labor. That bottleneck drives virtual screening: computational methods that sift through large libraries to identify molecules most likely to bind before expensive lab testing.

Virtual screening splits into two broad camps. Ligand-based approaches start from known active compounds and build models (such as pharmacophore features) that encode where key chemical interactions occur—hydrophobic regions, hydrogen bond acceptors/donors, and molecular shape—often using static constraints like exclusion volumes. Structure-based approaches instead assume a 3D receptor model is available and search over ligand structures to maximize predicted binding. However, conventional structure-based methods can struggle to reliably separate active from inactive ligands; an example involving thrombin ligands reportedly failed to distinguish high-affinity binders (low Ki) from poor binders, suggesting that more than standard docking-style signals may be needed.

Graph neural networks (GNNs) are presented as a newer, resource-efficient form of virtual screening that can improve accuracy and prediction speed, but only after careful parameterization of both ligand and receptor. Ligands are converted into molecular graphs where atoms become nodes with feature vectors (e.g., neighbor counts, hydrogens, formal charge) and bonds become edges with categorical or numeric descriptors (aromatic, conjugated, ring membership, and which atoms they connect). Receptors—proteins or polynucleotides—are handled by building graph representations from structural information. One common route uses adjacency matrices derived from 3D coordinates, where edges can be undirected (mirrored distances), directed (non-mirrored), or weighted (encoding bond/contact strength or distance). Other approaches avoid full coordinate dependence by predicting secondary structure first, then extracting contact maps and converting them into protein graphs with featurized amino-acid attributes like charge and functional groups.

Once ligand and receptor are both graphs, the GNN processes node and edge features through shared hidden layers, repeatedly aggregating information from neighboring nodes and edges up to a chosen depth. This recursive message passing builds graph-level embeddings in a fixed-size vector space, enabling a downstream dense layer to predict binding affinity for a ligand–receptor pair. Architecturally, the discussion contrasts recurrent-style GNNs (same weights reused until convergence) with convolutional-style GNNs (different weights per iteration).

The practical payoff is framed in early discovery terms: target interaction prediction is typically slow and expensive, yet computational screening can cut overall timelines and costs substantially. GNN-based affinity prediction is described as fast—milliseconds per ligand–receptor pair—and as extending docking capabilities by accepting receptor structures without coordinates, broadening what can be modeled when structural data is incomplete.