Fig. 1
From: Harnessing pre-trained models for accurate prediction of protein-ligand binding affinity
Overview. a Pre-training phase for ligands: Atom types are encoded via a linear layer as input to the transformer, while atomic coordinates are represented using a distance matrix and a spatial position matrix, creating an initial pair representation that serves as attention bias. Pre-training involves three self-supervised tasks: classification head, distance head, and coordinate head. b Affinity prediction for unknown protein-ligand complexes: Initial representations for proteins and ligands are derived from the pre-trained model and concatenated to form the initial representation of the complex, which lacks relative positional information between the protein and ligand. This representation is updated through transformer layers to predict binding affinity values. c Generation of spatial position representations from atomic coordinates: The spatial Cartesian coordinate system for a given atom is defined by its neighboring atoms (\(i-1\), i, \(i+1\)), allowing for the determination of the spatial positions of all other atoms