Co-folding, the future of docking – prediction of allosteric and orthosteric ligands

This is the authors' abstract. We don't add key points for this paper.

• Co-folding predicts protein-ligand complexes from sequence data. • Three co-folding methods were used to predict orthosteric and allosteric ligand binding poses. • Training biases favor orthosteric binding sites, posing a challenge for allosteric ligands. • Co-folding challenges: selection of the right ligand position and scoring. In drug discovery understanding protein structures is essential for comprehending their functions and interactions with drugs. Traditional methods like X-ray crystallography and cryo-electron microscopy have been used to solve these structures. Recently, computational biology has seen a breakthrough with deep learning algorithms capable of predicting protein structures based on amino acid sequences. These methods have now evolved into predicting protein-ligand interactions from sequence – co-folding methods. Despite the great advancement in the field during the last year, there are still open challenges. Here, we focus on the prediction of allosteric binding sites, using a dataset of 17 orthosteric/allosteric ligand sets. Three different co-folding methods – NeuralPLexer, RoseTTAFold All-Atom and Boltz-1/Boltz-1x – were used to predict both allosteric and orthosteric ligands. Using PoseBusters, the ligand quality was checked, with more than 90% of ligands predicted by Boltz-1x passing the default quality criteria. Boltz-1, NeuralPLexer and RoseTTAFold All-Atom still showing high quality drawbacks. The orthosteric ligands were well placed. However, instead of the allosteric pocket these deep learning approaches generally favor the orthosteric site, which is the one most represented in the training data.