How AI could fuel a mammoth antibiotic discovery

AI is speeding the discovery of much-needed new antibiotics and helping to identify promising antimicrobial compounds from some unusual sources – including the genomes of extinct organisms. Jasmin Fox-Skelly reports

It is more than 50 years since a new class of antibiotic effective against Gram-negative bacteria was discovered and brought to market [1]. Most available antibiotics, meanwhile, have a similar mode of action. The majority, including penicillin, contain a β-lactam ring in their structure. β-lactam antibiotics target both Gram-positive and Gram-negative bacteria, which are protected by a cell wall sandwiched between an outer and inner membrane. They work by inhibiting cell wall synthesis, but bacteria can easily evolve resistance to them by synthesising β-lactamase, an enzyme that attacks the β-lactam ring.

‘For years we’ve been stuck using older tools to discover antibiotics that are similar, if not in the same class, as antibiotics we already have,’ says James Collins, Professor of medical engineering and science at the Massachusetts Institute of Technology (MIT), US.

Yet there are hopes that this could soon change. AI promises to shake up the field, by slashing the time needed to screen potential drugs for efficacy. Rather than laboriously performing experiments by hand, computers can speedily rule out compounds that wouldn’t work based on what is known about their structure, or other data.

Normally it takes around three to six years to discover new clinical candidates, but with AI, scientists can now screen hundreds of thousands of preclinical candidates in a matter of hours.

‘Traditionally the way we’ve discovered new antibiotics is by going around nature, taking samples from water and soils, and painstakingly trying to purify and extract those compounds that might be active,’ says César de la Fuente, Presidential Assistant Professor at the University of Pennsylvania’s Perelman School of Medicine, US. ‘We have decades worth of biological information that people have sequenced. So, we can build algorithms to sort through all this data to try to find new compounds, which enables us to discover antibiotics at digital speed.’

Genomic approaches

In 2018, de la Fuente and colleagues used AI to design a new synthetic antibiotic, confirming that machines can help create useful molecules. Guavanin 2 kills bacteria by a novel mechanism of action, reducing infections in mice [2]. Since 2019, the lab has been building algorithms to trawl genetic databases looking for molecules with antimicrobial properties. Then, in 2020, the team searched the entire human proteome – every portion of DNA known to code for a protein – looking for antibiotic molecules [3].

‘We found thousands of new compounds in our own body with antibiotic properties that were previously unknown,’ says de la Fuente.

Next in 2023, they looked at the DNA of our extinct close relatives, the Neanderthals and the Denisovans. They found one peptide, neanderthalin-1, which, when recreated in a lab [4], was effective at treating bacterial infections in mice. The team went a step further in a study published earlier in 2024, which reported sampling the entire global microbiome – all the genomes of all the microbes sequenced to date, around 90,000 species [5].

‘We found almost one million potential new antibiotics encoded in all this microbial dark matter, which was amazing,’ says de la Fuente.

Finally in their most recent study, de la Fuente’s team used a new AI deep learning model called APEX to trawl the ‘extinctome’ – the 208 extinct organisms with sequenced genomes available to science [6]. They found tens of thousands of hits, and synthesised 69 of the most promising. These included peptides from the woolly mammoth, the extinct straight-toothed elephant, giant elk, and the mylodon, a cave-dwelling sloth-like creature that vanished around 12,000 years ago. Many of these compounds were effective at treating infections in mice, with the best activities comparable to the standard-of-care antibiotic polymyxin B.

These molecules, including mammuthusin, mylodonin, elephasin, megalocerin, and hydrodamin, now represent preclinical antibiotic candidates. They also have a completely novel mode of action, targeting the inner cell membrane.

‘Part of the molecule is positively charged, which electrostatically interferes with the negatively charged bacterial membrane,’ says de la Fuente. ‘Then at the same time, the hydrophobic building blocks that make up the majority of the peptide translocate into the membrane, poking a hole in it within a matter of minutes.’

Deep learning

At MIT, Collins is also hoping to use AI to discover new antibiotics, teaming up with Jonathan Stokes, Assistant Professor of biochemistry and biomedical sciences at McMaster University, Canada. In an initial proof of concept study in 2020, the researchers trained an AI algorithm to find chemicals that could inhibit the growth of E. coli. To obtain training data for the AI model, the researchers first exposed E. coli grown in a lab dish to thousands of different chemical compounds to see which ones could inhibit microbial growth. Then they taught the model the structure of each molecule, and whether it was effective at inhibiting bacterial growth. This allowed the algorithm to learn what structural features were most useful. Once the model was trained, the researchers used it to analyse 100m compounds the algorithm had not seen before – which it did in a matter of days. The results identified a compound, halicin, that was effective against E. coli and other bacteria resistant to antibiotics [7].

Halicin is an entirely new class of antibiotic, which appears to kill bacteria by disrupting their ability to maintain an electrochemical gradient across their cell membranes. ‘Whereas most antibiotics will target a specific protein, halicin acts on multiple proteins at once,’ says Collins. ‘It disrupts the bacteria’s ability to pump protons across its membrane, stopping it from generating energy.’

In 2023, the researchers used the same method to screen 7000 compounds to look for drugs that could prove effective against Acinetobacter baumannii, a species of bacteria often found in hospitals that can lead to pneumonia, meningitis, and other serious infections [8]. The screens identified nine antibiotics, including one, abaucin, which was extremely effective at killing A. baumannii. The researchers showed abaucin could treat wound infections caused by A. baumannii in mice and demonstrated its effectiveness against a variety of drug-resistant A. baumannii strains isolated from human patients. According to the scientists, the drug works by interfering with a process known as lipoprotein trafficking, which cells use to transport proteins from the interior of the cell to the outer cell membrane.

Although these two searches were successful, one limitation is that the algorithms act like a ‘black box’ – making it impossible to know which molecular features the AI is using to assess whether they make good antibiotics.

In a study in 2024, Collins and Stokes used a type of deep learning model that not only predicts antimicrobial activity, but also explains which structural features of the molecule likely account for that activity [9]. They used this model to look for compounds effective against methicillin-resistant Staphylococcus aureus (MRSA) and out of 12m molecules, found 280 promising candidates, which they then tested against MRSA grown in a lab dish.

The results identified two molecules – representing a completely novel class of antibiotic – that were highly effective against MRSA. In mice with MRSA skin infections and MRSA systemic infections, each drug was able to reduce the MRSA population by a factor of 10. Similar to penicillin, the new class discovered by Collins inhibits cell wall synthesis but does so via a very different pathway.

Resistance is futile

But will bacteria simply become resistant to these new classes of antimicrobials too? They may struggle to do so, at least at first, as many of the newly discovered compounds have multiple sites of action, so the bacteria would have to acquire many mutations rather than just one.

‘We think that bacteria are going to have a harder time developing resistance to these compounds compared with traditional antibiotics. But bacteria can evolve and mutate in a timescale of minutes. So probably over time they will develop resistance to anything that we throw at them,’ says de la Fuente.

Without new antibiotics, the World Health Organization (WHO) has suggested that 10m people could die due to drug resistant bacterial infections by the year 2050. A 2020 report by the WHO estimated that more than 50% of Klebsiella pneumoniae and Acinetobacter spp – bacteria responsible for common bloodstream infections in hospitals – are now resistant to mainstream antibiotics [10]. That number is 42% for E. coli, and 35% for MRSA, respectively. Bloodstream infections due to resistant E. coli and Salmonella infections increased by at least 15% in 2020 compared with rates in 2017.

The inevitability of bacteria evolving resistance is one problem preventing investment in the field. Another is financial. Clinicians will only prescribe new classes of antibiotics as a last resort – to safeguard them from developing resistance too – and so there is little financial incentive for pharmaceutical companies to develop new classes of drugs.

‘It costs roughly as much to develop an antibiotic as it does a cancer drug or a blood pressure drug, but antibiotics are sold quite cheaply,’ says Collins. ‘Also, if you come up with a really good antibiotic, the tendency is the community will shelve it and only use it when desperately needed, which poses challenges for a company because they won’t have any revenue.’

Without the prospect of making a profit, there is no hope of investment from pharmaceutical companies. So researchers must rely on funding grants from not-for-profits and public bodies. De la Fuente is considering how best to progress. ‘I’ve been working my whole career on antibiotic discovery. Our dream is to see some of these compounds helping people and benefiting humanity,’ he says.

‘Antimicrobial resistance is one of the biggest threats to humanity. If we don’t come up with new solutions, it’s going to be even worse. So that’s a huge motivation.’

Novel CRAB antibiotic

CRAB, also known as carbapenem-resistant Acinetobacter baumannii, causes blood, urinary tract, lung and wound infections. Flagged as an urgent threat in the US, mortality estimates for invasive CRAB infections range from 40 to 60%, in part due to the lack of effective treatment options.

Zosurabalpin is a completely new class of antibiotic, developed by researchers at Roche Pharma R&D in Basel, Switzerland. It kills CRAB using a novel attack strategy, by binding to and inhibiting a complex called LptB2FGC, responsible for transporting lipopolysaccharides from the inner to the outer membrane of the bacterium.

Lipopolysaccharides are an essential component of the outer membrane, without which the bacterium can’t function. So far, the drug has been shown to be effective at clearing infections in mice and rats and is currently being tested [11] in Phase 1 safety trials in humans. If licensed, it would be the first new class of antibiotics with activity against A. baumannii to be launched in over 50 years.

Zosurabalpin only works on A. baumannii because of the specific component of the machinery it targets. Unlike other broad-spectrum antibiotics, it won’t wipe out the gut microbiome, and the hope is there will be less selective pressure on the bacteria to develop resistance.

‘Drug resistance to all existing classes of antibiotics has been on the rise in various Gram-negative bacteria for several decades, but we have seen very few new antibiotics in development with the potential to overcome this threat,’ says Kenneth Bradley, Head of infectious diseases discovery at Roche. ‘Zosurabalpin has many features that position it to be a medical breakthrough, and future human clinical trials will inform whether it has potential to address a major gap in the fight against antimicrobial resistance.’