Somewhere in the genetic code of a woolly mammoth lies a compound that might kill drug-resistant bacteria. It’s called mammuthusin-2, and it’s one of several antibiotic candidates that César de la Fuente’s lab at the University of Pennsylvania has resurrected from extinct creatures using artificial intelligence.
The approach is called “molecular de-extinction,” and it represents one of the more creative attempts to solve a problem that kills 1.27 million people annually: antimicrobial resistance.
Mining the Extinctome
De la Fuente’s Machine Biology Group has built AI systems to scan published genetic sequences of species that no longer exist. Their targets include Neanderthals, Denisovans, woolly mammoths, giant sloths, and ancient zebras. The idea is simple: these organisms evolved antimicrobial peptides over millions of years to fight infections. Those genetic recipes are still encoded in their DNA, preserved in museum specimens and sequencing databases.
Their AI model, APEX, can identify thousands of potential antibiotic candidates in hours. Traditional drug discovery might take six years to reach the same point.
The names tell the story: mammuthusin (woolly mammoth), mylodonin (giant sloth), elephasin, megalocerin, hydrodamin. These aren’t theoretical compounds - they’ve been synthesized and tested.
Results So Far
The resurrected molecules work. In laboratory tests and preclinical mouse models, some show efficacy rivaling polymyxin B, a standard-of-care antibiotic typically reserved for serious infections when other options fail.
The lab has also turned to older life forms. In a study published in Nature Microbiology, they analyzed 233 Archaea species - Earth’s oldest microorganisms that survive in extreme environments like volcanic vents and toxic pools. From these ancient genomes, they identified over 12,000 potential antibiotic compounds, dubbed “archaeasins.”
Eighty were selected for initial testing. The logic: organisms that evolved under extreme stress likely developed potent chemical defenses.
ApexOracle: The Next Step
The lab is now developing ApexOracle, a multimodal AI system that goes beyond finding existing compounds. It can analyze a pathogen’s genome, identify genetic weaknesses, match those to antimicrobial peptides that might work, and predict how a designed antibiotic would perform in lab tests.
ApexOracle accepts three inputs: the pathogen’s genome sequence, a textual description of its traits, and a molecule. It outputs either a predicted efficacy or generates a new molecule entirely. In testing, it achieves state-of-the-art accuracy for predicting antimicrobial outcomes across both small molecules and peptides.
The goal is proactive drug design - creating antibiotics before specific resistant strains emerge, rather than scrambling to respond after they do.
The Race Against Resistance
Antimicrobial resistance is projected to cause 10 million deaths annually by 2050 if current trends continue. New antibiotic development has stalled for decades; pharmaceutical companies struggle to profit from drugs meant to be used sparingly.
De la Fuente’s approach sidesteps some barriers by radically accelerating the discovery phase. Finding candidates is no longer the bottleneck when AI can search millions of genetic sequences and identify promising molecules in hours.
The harder problems remain: clinical trials, regulatory approval, manufacturing, and the economic model that discourages antibiotic development. But having more candidates to work with is a start.
What This Means
The work demonstrates how AI can expand the search space for drug discovery into places humans wouldn’t think to look - or couldn’t search efficiently. Extinct organisms represent a vast, untapped library of molecular solutions evolved over millions of years.
It also shows AI’s role as augmentation rather than replacement. The system finds needles in haystacks; humans still need to validate the findings, synthesize the compounds, and run the trials.
Whether mammuthusin-2 or its relatives will ever treat human patients remains years away from knowing. But the approach has already yielded more antibiotic candidates than most pharmaceutical programs produce in a decade.