BY STEVE RANGER
Antibiotic resistance is one of the biggest health challenges facing the world. Now the combination of modern artificial intelligence (AI) tools and a collection of ancient and unusual organisms may be able to speed up research into new treatments.
Antibiotics become less effective as bacteria and viruses evolve, and existing drugs aren’t being replaced fast enough, contributing to the growing problem that could claim 39m lives globally by 2050.
Finding new sources of antibiotics can be a complicated, long and expensive process. That’s led researchers at the University of Pennsylvania’s Machine Biology Group to investigate the use of machine learning to seek out molecules with potential antibiotic properties, as well as looking at how generative AI could be used to help design new molecules.
This research has seen the team hunt through the human proteome but also those of our close relatives - Neanderthals and Denisovans, plus extinct creatures to find undiscovered antimicrobial sequences. The hope is that by identifying more candidate molecules more quickly, this can help speed up the delivery of new drugs.
‘It’s a whole gamut of AI systems that we’ve developed both for mining biological data, for designing things from scratch or evolving things that exist in nature and evolving them to make them better through computation,’ explains César de la Fuente, University of Pennsylvania Presidential Associate Professor and leader of the Machine Biology Group.
‘We have applied machine learning and AI in a very broad range to try to think outside of the box about antibiotic discovery and the big problem of resistance,’ he says.
‘Going way back I’ve always been fascinated by biology and its complexity and trying to understand nature and the world around us. I’ve always had the belief that if we were able to understand biology from first principles, we might be able to abstract that information and that knowledge to then apply it to solve many problems,’ he says.
‘I’ve always been fascinated by machines and the interplay between machines and human intelligence and how they can amplify human intelligence in some ways.’ AI is, for example, particularly good at processing huge amounts of data and finding patterns in very large data sets, while generative AI specialises in using existing data to create new outputs.
Over a decade ago, de la Fuente started to look at how machine learning could be applied to understand biology better and potentially develop with new antibiotics. ‘There was huge scepticism,’ he says because it was felt that biology was too complex to replicated in this way.
His initial projects looked at using an evolutionary computation approach to evolve molecules. This showed it was possible to evolve molecules that not only had the right structure but when synthesised could be effective. Since then, working at the University of Pennsylvania, de la Fuente’s team has developed deep learning algorithms that enable them to hunt through biological data from a wide range of sources to find potentially interesting molecules, and then test them for antimicrobial properties.
The fundamental idea behind the research is that, from a certain point of view, biology can be seen as a type of code.
‘DNA is a four-letter code, proteins and peptides are a 20-letter code and if you think about in just numbers, you can then [create] algorithms to mine all of that complexity, to try to find new molecules that might be hidden within,’ de la Fuente says.
When the team applied these algorithms to the human proteome - the proteins encoded in the genome - they found thousands of new molecules in the human body that were previously unrecognised as having a role as antibiotics or any role in the immune system. This lead them to propose the concept of ‘encrypted immunity’. These encrypted peptides serve as templates for antibiotic development, and may also play a role in host immunity.
‘We think there is a whole new branch of the immune system that is encrypted or hidden within our proteome that was previously unknown,’ he says.
Following this, the team decided to look at wider – and more unusual – sources of data.
‘We sampled all of ancient biology,’ de la Fuente explains. This was possible because, while the animals themselves have long since disappeared, their genetic information remain as a source of data to be analysed.
‘It enabled us to find new antibiotics in ancient penguins, magnolia trees that disappeared throughout evolution, in ancient seabirds, ancient elephants, woolly mammoths, giant sloths and many other creatures that used to roam around our planet at one point and they disappeared through the process of evolution.’
The work was published in Nature Biomedical Engineering in 2024; when mining the proteomes of all extinct organisms the team identified 11,035 sequences with broad-spectrum antimicrobial activity not found in extant organisms.
The team then synthesised 69 peptides to test their activity against bacterial pathogens. Most of these peptides killed bacteria by depolarising their cytoplasmic membrane, unlike better known antimicrobial peptides, which tend to target the outer membrane.
Compounds including mammuthusin-2 from the woolly mammoth, elephasin-2 from the straight-tusked elephant, hydrodamin-1 from the ancient sea cow showed anti-infective activity in mice. ‘Molecular de-extinction aided by deep learning may accelerate the discovery of therapeutic molecules,’ the paper said.
‘We think of a lot of those proteins and genes from ancient organisms as evolutionary relics or fossils, where we can extract a lot of useful information, so that’s what we are doing with this whole area we call molecular de-extinction,’ de la Fuente explains. ‘We see it as a conceptual framework to enable the discovery of new molecules and also to try to understand how they evolved and how that evolution influences how they operate in the world.’
The group has looked at other unexpected and underexplored sources of antibiotic potential. This sort of research is important because it can uncover unknown molecules with new physical and chemical attributes that in the future could be beneficial in creating new drugs, he argues, as well as helping to explain how peptides have evolved. One source looked at was venoms, which have evolved over millions of years to interact with a wide range of biological targets. In another paper published in 2024, the researchers explained that venoms are rich in bioactive peptides and proteins with pharmacological effects, including antibacterial activity. They investigated venom peptides using a sequence-to-function deep learning-model called APEX.
What makes venoms especially interesting is that, unlike most traditional antibiotics, which target specific bacterial enzymes or biosynthetic pathways, many venom peptides act by disrupting bacterial membranes, a strategy bacteria struggle to deal with. ‘Our findings support the notion that venom-derived peptides not only retain their antimicrobial function when extracted from their parent toxins but can serve as templates for future peptide-based therapeutics,’ they said.
Using the same techniques, the team has worked its way through sources covering the three domains of the Tree of Life, most recently by mining Archaea as a source of potential antibiotics.
In this study, published in Nature Microbiology in August 2025, the team used an updated version of APEX. By using a computational pipeline trained on known antimicrobial peptides, the tool identified another 12,623 potential antimicrobial peptides. Of these, 80 were synthesised and experimentally tested 80 ‘Archaeasins’, with 93% showing antimicrobial activity in vitro.
‘They are ancient organisms that can live in extreme environments like in volcanoes and places where a human would disintegrate within minutes. Every single antibiotic that has been described so far has come from fungi or bacteria but Archea have been completely ignored as a source of molecules so I think this is the first step towards trying to have more respect for Archea and to explore them and as a source of potential solutions,’ he says.
As well as using machine learning to identify existing molecules that could have an antibiotic capability, the researchers are also now looking at using generative AI to create new molecules.
The team detailed their work on AMP-Diffusion, which designs peptides using latent diffusion modeling. The team generated 50,000 candidate sequences, filtered and ranked them using the APEX deep learning model, and synthesised 46 top candidates which showed broad-spectrum antibacterial activity, against multidrug-resistant strains.
‘This study illustrates how generative AI can rapidly identify and optimise therapeutic peptides, offering a scalable and generalisable approach to antibiotic development. AMP-Diffusion sets the stage for future platforms that tailor peptides for specific pathogens or therapeutic targets,’ the paper published in Cell Biomaterials said.
More recently, the team published a study detailing a new tool de novo protein and peptide design, the Key-Cutting Machine. Unlike purely generative frameworks, which use trained models to rapidly generate large numbers of candidate sequences, this iteratively refines candidate sequences through structure prediction and optimisation. As a proof of concept, they used this new tool in antimicrobial peptide design by using a template antimicrobial peptide as the ‘key’, which they found yielded a candidate with ‘potent in vitro activity against multiple bacterial strains.’
Cutting the time to find pre-clinical candidates is one part of the puzzle when it comes to creating new antibiotics. And de la Fuente is positive about the impact that AI tools will have on broader scientific discovery. ‘I think it will have a huge impact; prior to the work that we’ve done, it took years to come up with pre-clinical candidates, a process that relied on trial and error, whereas now in minutes we can discover antibiotics. We can do science at digital speed,’ he says.
‘There is a lot of hype in AI to be honest, but the acceleration is real in scientific discovery, and I think in the field of antibiotics is an excellent example of something that has delivered,’ he says.
Visit Blog
