Database for extended ligand-target analyses (DELTA): a new balanced resource for AI applications in drug discovery

Abstract

We here present the DELTA resource, a database including balanced and annotated datasets of ligands for about 500 therapeutically relevant targets specifically collected for developing AI-based predictive models. For each target, DELTA comprises an optimized protein structure plus 200 experimentally tested ligands equally distributed between active and inactive molecules. All ligands are prepared by considering unspecified isomeric elements and combining semi-empirical calculations with MD simulations to explore their conformational space. The so-collected molecules allowed extended analyses of both ligands and targets, and the study presents some preliminary results. The performed analyses revealed that on average active ligands are larger than inactive molecules, while possessing a similar polarity. The scaffold analysis emphasized the expected and crucial role of aromatic systems, even though with some relevant differences between active and inactive molecules. Moreover, similar targets often show conserved binding sites and there is a limited but not negligible relationship between the similarity of binding sites and ligands suggesting that similar pockets tend to bind rather similar ligands. Finally, the collected biological data also allowed the analysis of the polypharmacological profile of the ligands endowed with more than one biological value. Most ligands bind two or three targets with diverse activities and almost always the bound targets belong to the same biological class. All the collected data are available for download at delta.unimi.it.