Scientists use AI to discover first new antibiotics in 60 years

Scientists at MIT and Harvard have just used deep learning to discover the first new class of antibiotics in over 60 years, and it takes aim at a deadly superbug.

The target is methicillin-resistant Staphylococcus aureus, better known as MRSA. This drug-resistant bacteria causes nearly 150,000 infections every year in the European Union alone, where antimicrobial resistance claims around 35,000 lives annually.

Peeking inside the AI black box

AI drug discovery often feels like an unpredictable black box. Models tell researchers that a molecule might work, but scientists are usually left guessing why.

A team of 21 researchers, publishing their findings in Nature, wanted to fix that. As lead author and MIT/Harvard postdoc Felix Wong explained, neural networks perform vast calculations mimicking brain connections, but rarely does anyone know what happens under the hood.

The researchers set out to open that black box. They trained an expanded deep learning model on roughly 39,000 compounds evaluated for antibiotic activity against MRSA, mapping explicit chemical structures to real-world results.

To ensure the predicted drugs would not harm humans, the team also trained three additional deep learning models to predict compound toxicity across three different human cell types. This allowed the system to balance powerful bacteria-killing power with minimal human cellular damage.

From 12 million molecules to two candidates

Armed with the new setup, the researchers screened roughly 12 million commercially available compounds. The platform flagged promising molecules from five distinct chemical classes.

The team then purchased around 280 of those compounds and tested them in the lab against live MRSA. Two compounds from the same structural class stood out.

In animal trials using two mouse models, one for skin infection and another for systemic infection, both antibiotic candidates successfully reduced MRSA bacterial loads by a factor of 10.

MIT professor and study co-author James Collins noted that the framework offers a level of time efficiency, resource efficiency, and chemical-structural insight that medicine simply has not had before.

Why it matters

The antibiotic pipeline has been practically dry for six decades while drug resistance quietly snowballs into a global crisis. What makes this breakthrough special is not just finding new molecules, but doing it with models that actually show their work by highlighting the specific chemical substructures responsible for antimicrobial activity.

We think the smartest strategic move here was integrating multi-cell human toxicity filtering directly into the computational screening phase. Historically, many promising drug candidates burn millions of dollars only to fail preclinical testing due to unforeseen cellular toxicity.

Still, it is worth keeping expectations grounded. Knocking MRSA populations down tenfold in mice is a massive milestone, but traversing human clinical trials, pharmacokinetic hurdles, and regulatory approval takes years. You cannot pick this up at a pharmacy anytime soon, but it proves generative and deep learning architectures can crack one of biology's hardest bottlenecks.