How to feed the machine: lessons from an AI antibiotics study
A machine learning study uncovering a broad-spectrum antibiotic is a case study in matching models to data
The recent AI-based discovery of broad-spectrum antibiotic candidates highlights the potential of new machine learning methods to find compounds humans might otherwise miss, but ups the ante on the quality and diversity of input data needed to pull that off.
In silico compound screening's promise of shrinking drug discovery timelines and costs while improving hit rates could particularly benefit antibiotic development, a field where new entrants have to compete against cheap generics while combating the threat of drug resistance.
Methods that are independent from human judgment could be especially advantageous for antibiotics discovery, since out-of-the-box scaffolds would be more likely to overcome resistance than those that resemble existing drugs.
"If you're asking a human to find something that has antibacterial activities, but doesn't look like antibiotics you're used to seeing, that would be a challenge," said Massachusetts Institute of Technology professor Jim Collins.
A February Cell paper argues