Researchers Used AI To Kill Drug-Resistant Bacteria

Two AI techniques allowed researchers to discover never-before-seen antibiotics that might neutralise two dangerous drug-resistant bacteria.

Artificial intelligence models did not create the new drugs on their own. Instead, the AI simply followed complex instructions to discover molecules that might be able to destroy Neisseria gonorrhoeae (gonorrhea) and Staphylococcus aureus (MRSA). 

The AI models generated millions of possible chemical compounds that would harm the bacteria and thus put a stop to infections. In each case, the researchers applied specific filters to narrow down the lists of compounds to adequate candidates. 

These filters included requirements that the resulting antibiotic should not harm humans nor share common traits with existing antibiotics that have lost their efficacy against the two bacteria. After applying these conditions, the researchers ended up with a few viable candidates that show promise in lab testing.

From millions of options, AI found a novel gonorrhea drug

To find a potential antibiotic for gonorrhea, researchers instructed the AI to create molecules based on a key bacteria-killing chemical fragment. They started with a set of 45 million fragments made up of all the possible combinations of 11 atoms and fragments from the Enamine Readily AccessibLe (REAL) space molecule repository.

From there, the AI refined the list to 4 million fragments that might kill the bacteria. After extracting chemical fragments that would harm the human body, researchers shrank the list to around 1 million candidates. After further tests, the MIT scientists ended up with a fragment called F1 that showed potential for addressing gonorrhea.

They fed the F1 candidate into two generative AI algorithms: chemically reasonable mutations (CReM) and fragment-based variational autoencoder (F-VAE). The former created molecules around F1 by modifying atom configurations and other characteristics.

The latter used learned patterns to forge complete molecules from a fragment. These two technologies produced 7 million potential candidates based on F1. That massive list ultimately shrank to some 1,000 viable compounds, out of which 80 were chosen for potential lab synthesis

 Just two of the 80 versions could be created, and only one (NG1) effectively destroyed gonorrhea in both a mouse model and lab dish.