Multi-Objective Evolutionary Design of Adenosine Receptor Ligands

A novel multiobjective evolutionary algorithm (MOEA) for de novo design was developed and applied to the discovery of new adenosine receptor antagonists. This method consists of several iterative cycles of structure generation, evaluation, and selection. We applied an evolutionary algorithm (the so-called Molecule Commander) to generate candidate A(1) adenosine receptor antagonists, which were evaluated against multiple criteria and objectives consisting of high (predicted) affinity and selectivity for the receptor, together with good ADMET properties. A pharmacophore model for the human A(1) adenosine receptor (hA(1)AR) was created to serve as an objective function for evolution. In addition, three support vector machine models based on molecular fingerprints were developed for the other adenosine receptor subtypes (hA(2A), hA(2B), and hA(3)) and applied as negative objective functions, to aim for selectivity. Structures with a higher evolutionary fitness with respect to ADMET and pharmacophore matching scores were selected as input for the next generation and thus developed toward overall fitter (better) compounds. We finally obtained a collection of 3946 unique compounds from which we derived chemical scaffolds. As a proof-of-principle, six of these templates were selected for actual synthesis and subsequently tested for activity toward all adenosine receptors subtypes. Interestingly, scaffolds 2 and 3 displayed low micromolar affinity for many of the adenosine receptor subtypes. To further investigate our evolutionary design method, we performed systematic modifications on scaffold 3. These modifications were guided by the substitution patterns as observed in the set of generated compounds that contained scaffold 3. We found that an increased affinity with appreciable selectivity for hA(1)AR over the other adenosine receptor subtypes was achieved through substitution of the scaffold; compound 3a had a K-i value of 280 nM with approximately 10-fold selectivity with respect to hA(2A)R, while 3g had a 1.6 mu M affinity for hA(1)AR with negligible affinity for the hA(2A), hA(2B), and hA(3) receptor subtypes.