Open the pod bay doors
Last week a group of scientists reported the discovery of a new antibiotic. They named the molecule ‘abaucin’ because it seems to kill the pathogen Acinetobacter baumannii, a bacterial species that can infect patients in hospitals and is frequently resistant to many other powerful antibiotics. The researchers say that abaucin works in a different way from existing antibiotics and seems to be a ‘narrow-spectrum’ antibiotic, targeting A. baumannii but having no effect against other bacteria such as E. coli.
A potential new treatment for a highly resistant pathogen is welcome news. Antibiotic resistance in A. baumannii has been a big problem in the Middle East, where the species has been well studied because of the number of difficult cases of wound infections it has caused in US soldiers since 2003. That’s produced a large literature written by American military clinicians on their struggles to treat these cases. (Reviewing a paper on the topic, I suggested the authors reconsider their description of the 2003 invasion of Iraq as ‘Operation Iraqi Freedom’.)
It has been reported that abaucin was ‘discovered using AI’. This needs a bit of unpacking. Finding any new drug means searching through ‘chemical space’ – the many possible configurations of atoms that can make up molecules. It’s difficult to get a grip on how vast this universe of possibility is. Most drugs consist of molecules with fewer than thirty atoms and a molecular mass of less than 500 daltons (a hydrogen atom has a mass of one dalton, give or take). It’s hard to estimate, but even if you restrict yourself to a handful of elements (carbon, hydrogen, oxygen, nitrogen, sulphur) there are at least 1060 possible molecules that fit these criteria. This is a big number, more than a thousand times the number of hydrogen atoms in the Sun. Exploring this chemical universe in its entirety is impossible. The hope is that using predictive algorithms from machine learning can help guide you to the right galaxy.
Getting to the right galaxy is hard – really hard. After a flurry of discovery in the middle of the 20th century, no new class of antibiotic has made it to market since 1987. (An antibiotic class is a group of molecules that have a common structure and work by a similar mechanism.) More recent antibiotics are derivatives, taking existing molecules and making a few tweaks to stave off resistance or improve their pharmaceutical properties – they are new stars in existing galaxies.
Part of the reason that finding a new antibiotic is even harder than finding other kinds of drug is that it must be, in effect, screened twice. In the early 20th century, Paul Ehrlich formulated the idea of a ‘magic bullet’, a chemical that affects only its intended target. The ideal antibiotic is a molecule that is both lethal for bacteria and has virtually no interaction with human cells: a double challenge. (Bleach kills bacteria, but drinking it would be a bad idea.)
Most antibiotics target fundamental cellular components that bacteria have but we don’t, thanks to billions of years of evolutionary separation. Penicillins, for example, target the proteins that build the bacterial cell wall which human cells don’t have. Alexander Fleming emphasised in his Nobel speech in 1945 that unlike other drugs his audience were familiar with, penicillin was ‘to all intents and purposes non-poisonous’ to humans.
Ultimately, the only way to find out what will really happen when your new molecule interacts with the human body is to put it inside a person: however thoroughly you’ve tested it in the lab, it may be transmogrified quite unexpectedly. The long timescales and huge costs involved in getting an initially promising molecule through the various stages of animal and human safety trials, even before you get to efficacy trials, mean that it’s tempting only to look at molecules that have been previously approved for human use or used in pre-clinical trials. Projects such as the Drug Repurposing Hub, used by the abaucin team, collate thousands of already approved drugs and potential candidates. As Stephen Fry once wrote: ‘An original idea. That can’t be too hard – the library must be full of them.’
The researchers ‘screened 7500 molecules for those that inhibited the growth of A. baumannii in vitro’, then ‘trained a neural network with this growth inhibition dataset and performed in silico predictions for structurally new molecules with activity against A. baumannii. Through this approach, we discovered abaucin.’ It isn’t a new chemical, though: it was in the database because it is known to bind to a particular human cell receptor, which might treat a certain complication of diabetes. In practice, giving it as an antibiotic would mean trading off those human effects against the bacterial ones – by construction, it cannot be a magic bullet. Presumably for that reason, the researchers focused on topical (rather than systemic) treatments: they found that abaucin reduced the amount of infection by A. baumannii in wounds on the backs of anaesthetised mice. Not putting the antibiotic into the body circumvents the problems that can arise if you give it orally or intraveneously, but also limits its application.
Abaucin isn’t the first potential new antibiotic to have been found with the help of AI. Three years ago, some of the same researchers reported the discovery of a molecule in the Drug Repurposing Hub that they (re)named halicin after the sentient computer in 2001: A Space Odyssey. In that case, some experts expressed scepticism that the machine learning techniques would really shorten the overall development cycle of drugs. The discovery of abaucin shows that AI is helpful for the early stage of winnowing down the vast space of chemical possibility, but there’s still a lot to do from that point onwards. One commentator believes that we don’t currently have the data to use AI to find molecules more likely to succeed in clinical trials: ‘Someday we may well. But that day is not today.’
As well as powerful neural networks, the machine learning model depends on the existence of carefully collected data from thousands of experiments. It’s still a vast screening project, just not as vast as it would be without the AI component: it uses the data to find the best ‘ready to use’ molecule from the available options. There have been high hopes for drug design using computers since the 1980s, but as the drug developer Lynn Silver has pointed out, most of the antibiotics we rely on weren’t products of ‘rational’ design but rather the laborious and ‘irrational’ process of screening chemicals in the lab: ‘Not especially innovative, but it worked.’