Data set composition.

Composition of the data sets generated. The allosteric set was obtained via text mining of abstracts; the non-allosteric (Full) set was the remainder of ChEMBL obtained using the same constraints as the allosteric set (e.g. limit bioactivity to primary assay). The non-allosteric (balanced) set was derived from the non-allosteric (full) set by taking a random percentage of each L2 target class present in the allosteric set. The classes ‘Organic’ and ‘Inorganic’ were subsets of the ‘Small molecules’ class. ‘Peptide’ was a subset of the ‘Biologicals’ class. Abbreviations: L1 – Level 1 target classification, L2 – Level 2 target classification.