A drug library can hold billions of molecules; a target has one binding site. Today’s digest is about teaching machines to guess, fast, which molecules will stick — the scoring engine of virtual screening. Ragoza et al. showed a CNN reading ‘a comprehensive three-dimensional representation of a protein-ligand interaction’ can beat AutoDock Vina ‘both for pose prediction and virtual screening.’ Jiménez et al.’s KDEEP predicts absolute affinity with ’each prediction taking a fraction of a second’ — while warning ‘accuracy is still very sensitive to the specific protein used.’ Stępniewska-Dziubińska et al.’s Pafnucy grids the complex and treats ’the atoms of both proteins and ligands in the same manner.’ Zheng et al.’s OnionNet stays robust on docked, not crystal, poses. And when there’s no structure at all, Öztürk et al.’s DeepDTA scores affinity ‘using only sequence information,’ while Nguyen et al.’s GraphDTA represents ‘drugs as graphs.’ Ranking the haystack.
