The drugs of the future are in the animals of the past

César de la Fuente’s lab has a knack for finding antibiotics in unusual places. He doesn't trudge through swamps or remote forests like a pharmacological Indiana Jones. His lab instead combs through genetic data collected from creatures across all time.  

In just the last few years, they’ve documented unreported antimicrobial compounds hidden in the genomes of Neanderthals, the world’s microbes, and within ourselves. Now, their latest feat carries the torch thanks to a brand new machine learning algorithm they call APEX: antibiotic peptide de-extinction. 

APEX is sort of like a drug-sniffing dog. But instead of a German Shepherd named Bruce scanning carry-ons for bags that smell like weed, the AI scours extinct animals’ genetic sequences for patterns it has learned to mean “this gene makes germ-killing molecules.”

Before this project, de la Fuente’s team at the University of Pennsylvania had only explored the genetic data of a few animals at a time. They dreamed bigger. “Here, we challenged the team to see if we could explore every extinct organism known to science,” de la Fuente said. “So we had to develop a more powerful AI model.”

What they’ve found are peptides, short chains of amino acids that, in this case, can infiltrate and destroy pathogenic bacteria, just like an antibiotic. When APEX alerted them to promising sequences, the researchers could cook up that molecule and try it out in test tubes and mice infected with the pathogen Acinetobacter baumannii. 

The AI speeds up candidate discovery — a point of importance that de la Fuente stresses repeatedly. “Antimicrobial resistance is one of the greatest threats to humanity,” he said. Microbes are evolving resistance to our drugs faster than we can replace them. It takes an estimated 15 years and $1+ billion to develop each new antibiotic. Traditionally, a grad student might wander around collecting environmental water and soil samples to purify in the lab, searching for a hit. It’d take years and an entire PhD to painstakingly gamble on this approach. 

“This dramatic acceleration of discovery has been amazing,” de la Fuente said. “We can come in the morning, have a cup of coffee, and by dinnertime the computer has already given us hundreds of thousands of new sequences of molecules to explore.” 

Not every one of those thousands is a winner. Just like narcs tipped off by drug-sniffing dogs, the team investigates each alert, refining the list of predicted candidates down to a testable few: from 37,176 germ-killing peptides across extinct species in their available databases down to 69 peptides chosen test in their lab, according to their study published in Nature Biomedical Engineering.

“We can come in the morning, have a cup of coffee, and by dinnertime the computer has already given us hundreds of thousands of new sequences of molecules to explore.” 

“We're literally reviving these molecules,” de la Fuente said. Over 11,000 of the discovered peptides no longer exist in nature. The five leading candidates are molecules from the wooly mammoth (peptide mammuthusin-2), straight-tusked elephant (elephasin-2), Steller's sea cow (hydrodamin-1), giant sloth (mylodonin-2), and giant Irish elk (megalocerin-1) — all extinct. Each stand-out obliterated deadly pathogens in test tubes, and treated mice infected with A. baumannii as well as the standard doctor-prescribed antibiotic polymyxin B did. 

“If you had asked me six years ago whether this would have been possible today, I would have probably predicted that we would have needed five more years to get to this point,” de la Fuente said.

We’ve clearly entered a drug discovery era that’s vastly different. I earned my PhD by developing new antibiotics so — if you’ll allow me to editorialize a little right here — this is pretty damn cool. Still, I must emphasize: we’re not mice (“rodent men” aside) so a study like this can only say so much about how a sea cow’s long-extinct peptides can rescue people from superbugs. 

Also, as this type of lightning-quick drug discovery amasses piles of potential drugs, I’m left wondering: what happens from here? It’s all but certain that even the five best de-extinction peptides won’t all get clinical trials in humans soon. How can we modernize R&D pipeline after finding cool molecules so that we can pick out the best even faster? 

De la Fuente’s reflex is to be optimistic: “If we flood the pipeline with hundreds or thousands or millions of candidates, the likelihood of at least one of them making it through is a lot higher,” he said. “But we may need to build additional AI models to help us filter all those preclinical candidates.” 

So a modernized R&D could mean more AI from start to finish. To accomplish that though, de la Fuente needs better data. Data that shows how stable these peptides are in the chaotic environment of human bodies; data on which peptides can work in combination with conventional antibiotics to stave off drug-resistance; data about toxicity to human cells and immune response; and data that helps scientists predict which sequences would translate best from mice to people. “The goal in the future is to create a multimodal super algorithm that can take into account all those different factors at once,” he said.

But getting game-changing data will require a lot of human work and ingenuity. “That's the other thing: Everybody's excited about AI but people have not really caught up to the fact that AI is just a tool.” Even in a future of AI drug discovery, human researchers will be far from extinct.