Ancajas, C. M. F.; Shuster, I. E.; Walker, A.

One of the major challenges in natural product discovery is the prioritization of compounds with useful activities from microbial sources. In particular, this is a challenge in genome mining for novel natural products, where the structures and activities of compounds produced by bioinformatically identified and uncharacterized biosynthetic gene clusters remain unknown. Here, we utilize a machine learning model to predict the antibacterial activity of a natural product from its biosynthetic gene cluster (BGC). We prioritized the strain Amycolatopsis azurea DSM 43854 which was predicted by machine learning to have the capacity to produce multiple natural products with antibacterial activity. Together with bioactivity-guided fractionation, we isolated dipyrimicins A and B from Amycolatopsis azurea DSM 43854 and, for the first time, linked them to their BGC. This dip BGC was predicted by our model to encode a product with 75% antibacterial probability and shares only 40-52% similarity with previously characterized BGCs. We confirmed the antimicrobial properties of the dipyrimicins against a few test strains and identified key tailoring enzymes, including an O-methyltransferase and amidotransferase, that differentiated them from other related 2,2\'-bipyridine biosynthetic pathways. Importantly, As the dip BGC was not in the training set of the model, our results demonstrate the ability of the model to generalize beyond its training set and the potential of machine learning to accelerate novel bioactive natural product discovery and deorphanization of biosynthetic gene clusters.