De novo drug design is a long-standing challenge in computational chemistry. Much recent work has centered on using generative neural networks and related technologies to create chemical libraries that reproduce property or structural distributions that mimic those of compounds known to be active against a selected target. The AI-driven Drug Design module in ADMET Predictor® takes a different approach: in AIDD, one or more molecules are submitted to randomly chosen SMIRKS transformations, yielding products that are then evaluated against a panel of structural SMARTS filters and objective functions. Those that pass the filters and are Pareto optimal with respect to the objectives and to the prior generation of analogs are carried forward as the next generation, which is itself subject to additional transformations, and so on.

The case study to be presented here involves a series of triazolopyrimidines (TzPs) that inhibit the dihydroorotate dehydrogenase of the malarial parasite Plasmodium falciparum – i.e., PfDHODH. Examples from very early in the synthesis program were run through AIDD using four objective functions: a general artificial neural net ensemble (ANNE) model of PfDHODH inhibition with a single hidden layer; bioavailability estimates obtained from high-throughput pharmacokinetic (HTPK) simulation; synthetic difficulty estimates; and a subpanel of absorption, distribution, metabolism, excretion and toxicity (ADMET) property estimates encoded as series of rules (AIDD Risk).

The output analogs obtained afforded good structural coverage of the activity space, were novel but synthetically reasonable, and included several compounds synthesized later in the TzP program. Those subsequently synthesized analogs were not involved or represented in the data set used to build the ANNE activity model. Hence, despite the data set involved being technically retrospective, this case study is largely prospective in nature.

Marvin Waldman

Appearances