Identifying protein-ligand binding sites is crucial for drug discovery. Current models are often developed and evaluated on small structures from high-resolution databases. When they're applied to larger protein complexes, they often do worse. Cryo-EM (electron microscopy) is getting more popular as a structure imaging tool because it can capture larger protein complexes but at lower resolution.
This study curates a novel database (EMD-Ligand Dataset) consisting of Cryo-EM maps and ligand-binding partner information. State-of-the-art protein ligand binding site prediction (PLBP) methods, while effective on conventional databases, perform poorly on the EMD-Ligand Dataset, emphasizing the need for further research.
The study proposes DeepTracer, an end-to-end Cryo-EM protein structure modeling method. It predicts ligand binding sites directly from Cryo-EM maps; they demonstrate its effectiveness by generating drug-like ligands from predicted binding pockets.