SoylentNews

upstart writes in with an IRC submission for AzumaHazuki:

IBM develops AI to invent new antibiotics – and it's made two already:

The IBM Research team created an AI system that's much faster at exploring the entire possibility space for molecular configurations. First, the researchers started with a model called a deep generative autoencoder, which essentially examines a range of peptide sequences, captures important information about their function and the molecules that make them up, and looks for similarities to other peptides.

Next, a system called Controlled Latent attribute Space Sampling (CLaSS) is applied. This system uses the data gathered and generates new peptide molecules with specific, desired properties. In this case, that's antimicrobial effectiveness.

But of course, the ability to kill bacteria isn't the only requirement for an antibiotic – it also needs to be safe for human use, and ideally work across a range of classes of bacteria. So the AI-generated molecules are then run through deep learning classifiers to weed out ineffective or toxic combinations.

Over the course of 48 days, the AI system identified, synthesized and experimented with 20 new antibiotic peptide candidates. Two of them in particular turned out to be particularly promising – they were highly potent against a range of bacteria from the two main classes (Gram-positive and Gram-negative), by punching holes in the bugs' outer membranes. In cell cultures and mouse tests, they also had low toxicity, and seemed very unlikely to lead to further drug resistance in E. coli.

Journal Reference:
Payel Das, Tom Sercu, Kahini Wadhawan, et al. Accelerated antimicrobial discovery via deep generative models and molecular dynamics simulations [open], Nature Biomedical Engineering (DOI: 10.1038/s41551-021-00689-x)

Original Submission