Designing synthetically accessible natural product mimetics by machine learning

The structural complexity of natural products often requires scaffold hopping to synthetically more easily accessible isofunctional compounds. We present a computational concept for (i) ‘de-orphaning’ natural products by target prediction, (ii) scaffold-hopping from pharmacologically active natural products to synthetically accessible compounds by holistic similarity searching, and (iii) the automated de novo design of natural-product mimetics. In a first prospective application, we used natural cannabinoids as queries in a chemical database search for novel synthetic modulators of human cannabinoid receptors. Of the synthetic compounds selected by the new method, 35% were experimentally confirmed as active, with nano- to micromolar in vitro activities. These cannabinoid receptor modulators were structurally less complex than their respective natural product templates. In a second prospective study, novel retinoid X receptor (RXR)-targeting natural products were computationally identified. These natural products then served as templates for ligand-based de novo design. The computer-generated mimetics feature innovative, synthetically easily accessible molecular scaffolds and inherited the biological activities of their natural templates. In a third application, we synthesized the top-ranking compounds designed by a deep learning model. These compounds revealed potent bioactivities. Apparently, the ‘artificial intelligence’ intrinsically captured relevant chemical and biological knowledge of both natural products and synthetic compounds without the need for explicit design rules