Key Points
- New antibiotics are needed to combat drug-resistant bacteria; artificial intelligence (AI) helps accelerate the search.
- Machine learning algorithms can parse through the biological blueprints of living and extinct organisms to uncover molecules with antibiotic potential.
- Researchers are also using generative AI to design "new-to-nature" antibiotic molecules from scratch.
- While AI can help discover fresh antibiotic candidates, their development and clinical deployment remains a challenge.
On a bench in a Philadelphia lab, a robot the size of a microwave clicks through tiny vials, building molecules that existed only as lines of code a week earlier. Some of those molecules trace their lineage to woolly mammoths and Neanderthals-not their bones, but their biological blueprints (e.g., genomes and proteomes). Others were conjured up from scratch by algorithms. All are being tested against bacteria that are increasingly shrugging off our best medicines.
Antimicrobial resistance (AMR) already kills large numbers of people each year worldwide and threatens many more in the decades ahead. Yet, the world has not discovered a new class of antibiotics in decades, and the pipeline of drug candidates remains sparse. This is why a growing number of scientists are turning to artificial intelligence (AI) to compress the long, failure-prone search for antibiotics into something faster, cheaper and broader, leveraging digital tools to uncover or design novel candidates.
"A number of years ago, we had this idea of thinking of all of biology as an information source, as a bunch of code," said César de la Fuente, Ph.D., a Presidential Associate Professor at the University of Pennsylvania. "If you think about it that way, you can devise algorithms to sort through that code and identify things that might look like antibiotics."
But while AI could be a game-changer for antibiotic discovery, it is not completely re-writing the rules. Turning an antibiotic candidate into a viable therapeutic depends not only on computers, but also on mobilizing resources that help transform drug contenders into clinical tools.
Discovering New Antibiotics is Hard
Though much has happened over the last 40 years, discovery of a new class of antibiotics is not one of them. This isn't great, considering AMR is expected to kill 39 million people over the next 25 years (it already kills roughly 1 person every 20 seconds). There are antibiotics in pre-clinical and clinical development. However, according to the World Health Organization (WHO), there are not enough innovative compounds in the pipeline to address the escalating AMR threat.
The dearth in antibiotic discovery is not for a lack of effort. Most of the world's existing antibiotics were discovered by prospecting in nature for compounds that killed microbes without harming people. This traditional approach worked-until it didn't. "Its a very physical process where scientists literally go around nature and dig into dirt and water samples and try to purify active compounds that might be contained within all of that complex organic matter," de la Fuente said. "But, as you can imagine, that's a painstaking process that is unpredictable."
Complicating matters, antibiotics can't simply destroy bacteria outright; the host must also be taken into consideration. Meaning, effective antibiotics must also be soluble, get to the right anatomic site at the right concentration and not kill the host in the process, said Jonathan Stokes, Ph.D., an assistant professor of biochemistry and biomedical sciences at McMaster University. "Given that drug discovery for any therapeutic application is a multi-property optimization problem, it necessitates that we look at a ton of molecules."
High-throughput screens, in which hundreds of thousands of compounds are experimentally tested against bacteria/purified microbial targets, have helped streamline drug discovery. Yet, they are also costly, time-consuming and tend to bias toward compounds that can't surpass the outer membrane of gram-negative bacteria (some of the trickiest AMR organisms to combat).
Confronted with these barriers, researchers are increasingly turning toward something that, unlike the world's repertoire of antibiotics, has rapidly advanced: AI.
AI Accelerates Antibiotic Discovery
When we talk about AI, we are usually referring to machine learning (ML)-a subset of AI in which algorithms learn from training data to make predictions about new data; the more quality data going in, the better the predictions coming out. ML models can quickly identify patterns and distill countless possibilities into a smaller number to guide decision-making. Such technology has dramatically accelerated the discovery of new antibiotic candidates.
For instance, scientists can feed an ML model the chemical structures of thousands of compounds experimentally demonstrated to be either active or inactive against a target bacterium. When presented with billions of new chemical structures, the model parses out potential "hits" based on what it learned differentiates active compounds from inactive compounds. "ML models are suggestion generation boxes," Stokes said. "I, as the scientist, can then take those suggestions, pick the best ones and move into the lab and run the real-world experiment.
The data underlying these suggestions are vast, varied and largely underexplored. The efficiency and speed of ML means scientists can now look for new antibiotics in places they never have before-including the distant past.
Mining (Ancient) Biology for New Antibiotics
Biology has long been a primary source for new antibiotics (think soil microbes). However, in this era of computational power, finding drug candidates is less about digging through dirt, and more about digging through data.
De la Fuente's lab, which has become a hub for the emerging field of AI-first antibiotic discovery, spearheaded the approach of using ML to comb through genomic and proteomic sequencing data spanning the Tree of Life, pulling out snippets that encode products with antimicrobial potential. The lab's primary interest is in identifying antimicrobial peptides-short amino acid sequences with extensive diversity and attractive antibiotic properties-from the blueprints of both living (e.g., humans) and extinct organisms.
Case in point: his team built a model that, upon parsing through the proteomes of our closest extinct relatives, Neanderthals and Denisovans, uncovered peptides with predicted antimicrobial activity. When synthesized in the lab, the peptides effectively killed the pathogen Acinetobacter baumannii in vitro and in vivo.
With another model, they mined the proteomes of the woolly mammoth, straight-tusked elephant, giant sloth, ancient sea cow and other archaic animals. Several peptides exhibited anti-infective activity in mice with skin abscess or thigh infections. Notably, the peptides (with names like mammothisin-1 and elephasin-2) were generally as effective as polymyxin B, an existing peptide antibiotic, and killed bacteria by depolarizing their cytoplasmic membrane.
De la Fuente noted this process of so-called molecular de-extinction can provide important insights into biology and evolution (i.e., how do changes in a sequence influence molecular function over time?). At the same time, it unlocks molecules that may have benefited long-dead organisms, and could also help us address today's problems, including AMR. "Evolution encodes immense biological intelligence," he said. "We see molecules as documents of that history-molecular fossils we can read to extract useful insights and, potentially, molecules that could help humanity."
Generating New Antibiotics Using AI
But what if the most promising antibiotics are molecules that don't exist-and have never existed? The number of theoretically possible chemical compounds is roughly 100 times that of all the grains of sand on Earth. "It's such a vast number, it's nonsensical," Stokes said. Some compounds may be buried in biological code; others are represented in online repositories housing millions of chemical structures, which ML can sift through for antibiotic candidates. However, considering the sheer number of chemical possibilities, such repositories are limited.
The solution: design new compounds from scratch. Using what's known as generative modeling, scientists train ML models on molecules known to be active or inactive antibiotics. "But, instead of now showing pictures of new molecules from the internet, you say, 'Hey, model, draw me a brand-new picture of a molecule that you think is going to be active,'" explained Stokes, highlighting that this dramatically opens the search space for novel chemicals.
Generative models can also unify both biology and chemistry. De la Fuente's group recently described a system that takes a pathogen's genome sequence-or even a brief phenotypic sketch-and suggests "new-to-nature" molecules to neutralize it, a strategy that could be deployed against emerging threats. The team has also built generative AI tools to systematically tune and enhance the antibacterial potency of candidate compounds.
What's tricky about generative AI is that it tends to invent compounds that seem great in the digital world but are next to impossible to synthesize in the real one. And if you can't make a compound, you can't use it. It's a roadblock that Stokes and his collaborators have worked to overcome.
They built a generative ML model that pulls from libraries of multi-atomic molecule "building blocks," rather than piecing together molecules atom-by-atom, as most generative models do. Because how each building block reacts with every other fragment is known-and can be feasibly, quickly and cheaply made using standard chemical reactions-the model's output molecules are not just theoretically promising, but synthetically tractable. Indeed, compounds designed by the model demonstrated antibacterial activity against A. baumannii and other pathogens in the lab.
"Now we've no longer said, 'Hey, model go nuts,' because it's going to draw something crazy. We can constrain it to this chemical space," Stokes explained. The model still generates a cornucopia of new chemicals-46 billion in its current form-but ensures their experimental, and, thus, therapeutic, possibility.
It Takes More Than AI to Make an Antibiotic
Armed with AI's affinity for efficient data-mining and molecular design, there is no doubt we'll see some new antibiotics on the market soon, right?
Not quite. Discovering lead compounds is a small step in the yearslong process of drug development. Many hindrances occur after the discovery phase, when candidates fail in clinical trials for various reasons (e.g., toxicity), or don't have the financial backing to progress through the pipeline. It takes a lot of money to develop an antibiotic. But because they cost little after commercialization, and people only take them for short periods of time, pharmaceutical companies see little, if any, profit. For this reason, both Stokes and de la Fuente emphasized that the success of new antibiotics-regardless of AI's involvement-will rely on governments and philanthropists putting up the funds as a service to public health.
When it comes to AI-assisted antibiotic discovery, the future depends on data, data and more data. ML model predictions are only as good as the data they are trained on, meaning developing quality, standardized, biologically relevant training data sets is important. Anticipating that need, de la Fuente's lab spent years assembling rigorously curated training data for its models. The team measured minimum inhibitory concentrations (MICs) for thousands of molecules across diverse bacterial strains, holding temperature, pH, media and other variables constant so results would be comparable. It's painstaking work, but, as de la Fuente argues, standardization is what turns clever code into models that are genuinely useful and meaningful.
Those models could eventually do more than find or generate antibiotic structures. Stokes noted that researchers must ultimately think beyond using AI to uncover drug "hits" and begin integrating it into the entire antibiotic development process, from preclinical studies to clinical trials. For example, models could help predict how likely a drug is to succeed (or fail) in clinical trials and why.
Nevertheless, while there is a vast amount of potential, algorithms won't defeat AMR on their own. "AI is just a pipette," Stokes said. "It's just another tool in our toolbox to accelerate solutions to the problems we were going to try to address anyway. That's it-no more and no less than that."
AI can help yield fresh antibiotics with a fighting chance at success. But, when all is said and done, getting antibiotics out of the lab, through the development pipeline and ultimately to the people who need them, remains a human endeavor.
In addition to antibiotic discovery, AI is being used to detect antibiotic resistance in bacteria. Learn how scientists are leveraging AI to develop detection tactics that are easy, accurate-and quick.
