A headline caught my eye last week… well other than the normal coronavirus shenanigans anyway; “A powerful antibiotic that kills some of the most dangerous drug-resistant bacteria in the world has been discovered using artificial intelligence”.
Ignoring the sensationalised headlines I dug out the research paper, published in Cell last week, and had a look at what was going on, obviously it was a quiet day on the coronavirus front!
It can take many years to identify potential new compounds with antibiotic activity (or disappointingly “rediscover” an already existing compound) and added to this the researchers have to start the long process of experimental and clinical trials to see if the new antibiotic is safe as well as effective. Many drugs fail at this later hurdle because they are just too toxic to use in humans, so all of that early research work and time has been wasted. Essentially the cost of developing a new antibiotic is said to be too prohibitive for drug companies, especially as treatment courses for infections are so short compared with life long, lifestyle adapting drugs such as statins, beta-blockers and proton pump inhibitors, etc.
In order to make this process more cost effective many companies look at antibiotics that already exist and try to modify the structures in order to produce a more active compound. It’s kind of cheating! And the result is a lot of “generations of…” cephalosporins, fluoroquinolones, aminoglycosides etc. The catch is that often these “new” antibiotics are susceptible to the same mechanisms of resistance as the original antibiotics, so when resistance occurs (“bacteria mutate and evolve to sidestep the mechanisms that antimicrobial drugs use to kill them” The Guardian) it takes out the whole class; making the old and the newly “discovered” antibiotic useless.
So how does AI make the process faster and cheaper?
Well those clever bioengineers, computer scientists, artificial intelligence (AI) specialists and biomedical scientists (what a great team!) used AI to identify “proper” new antibiotics…and quickly!
Screening for new antibiotics
The team from the USA developed an AI system (essentially a very complex computer program) that they “taught” how to recognise potential new antibiotics. The teaching was based around known compounds shown to have antimicrobial activity against E. coli or others known to not have activity. The original set of compounds consisted of 120 known “antimicrobials” and 2,200 other compounds.
Once the AI had been taught, it was given 6,111 molecules from the Drug Repurposing Hub (DRH) to analyse further. The DRH is a database of all sorts of drugs at various stages of research for use in human disease. The AI identified 51 molecules predicted to have good activity against E. coli. Further prioritisation based on factors like more advanced phase of clinical study, lack of similarity to other antibiotics and reduced predicted toxicity, identified 2 specific compounds which they studied further.
Using the factors “advanced stage of study” means more is known about the drug being studied, added to this the “reduced toxicity profile” and the “lack of similarity to other antibiotics” in the AI’s predictions means the new compound should “really be new” and the risk of resistance to it should be low. The use of AI to rapidly screen potential antibiotics like this can potentially save years in terms of new drug discovery, thereby directing the dwindling (or lack of) resources for new antibiotic discovery to better “targets” for further investigation.
The team identified a compound, currently being studied for the treatment of diabetes, to study further, its name was “c-Jun N terminal kinase inhibitor SU3327”… but don’t worry, I’m sure a marketing agency was called in as it’s now called “Halicin” (apparently named after the rogue computer Hal 9000 in 2001 A Space Odyssey)!
So this was a diabetes treatment, not an antibiotic? With no known antimicrobial properties at all… yep… cool huh!?
How good is Halicin?
So having predicted Halicin should be a good antibiotic, how good actually is it?
Halicin was shown to be effective against E. coli at an MIC (Minimum Inhibitory Concentration) of 2mg/L; this about the same as Meropenem and Gentamicin and more active than a lot of other antibiotics we currently use such as Amoxicillin, Piptazobactam and some of the cephalosporins.
Halicin was also shown to have good activity in vitro against antibiotic resistant E. coli carrying genes for resistance to Colistin, beta-lactams, aminoglycosides, carbapenems and also multidrug resistant (MDR) Acinetobacter baumanii. Sadly it wasn’t active against Pseudomonas aeruginosa, but you can’t have everything.
The USA researchers also went on to assess the activity of Halicin in a mouse model of infection with MDR A. baumanii which showed it to be very effective and well tolerated.
How does Halicin work?
Halicin acts by disrupting proton motive force preventing the bacterium producing its own energy. This is a novel mechanism of action and makes Halicin bactericidal i.e. it actually kills bacteria.
Given that the structure of Halicin is different to other antibiotics and that it has a novel mechanism of action it is unlikely that pre-existing resistance mechanisms within bacteria will occur to Halicin. This has so far been found to be true; in fact the researchers were unable to make a Halicin resistant E. coli using conventional laboratory methods for doing so e.g. exposing bacteria to slowly increasing concentrations of antibiotic to select resistant mutants.
Has the AI system identified any other new antibiotics?
Rather than sitting back on their laurels and basking in the adoration of the scientific community, the researchers in the USA have gone even further. They have looked at even bigger databases of chemical compounds to find more antibiotics.
The AI was used to look at 107 million promising compounds from the ZINC15 database (a virtual collection of about 1.5 billion molecules). From this piece of work they identified 23 compounds worthy of further study. Of these 23 compounds 2 were shown to have broad-spectrum activity in the same way as Halicin but were novel compounds in terms of structure. The researchers have called these new antibiotics ZINC0001000032716 and ZINC000225434673… maybe there’s also some work needed to give them catchier names….
So there you have it, AI has found 3 completely new antibiotic compounds which may have a use in the treatment of antibiotic resistant infections in the future… amazing!
Is the antibiotic crisis over?
It’s a promising change in the way new antibiotics could be discovered and it could significantly cut the time and cost involved in producing new drugs making them more likely to be effective, and better expected to pass safety studies and clinical trials. However before you start singing and dancing in the streets crying that the post-antibiotic era has been averted just think about it a bit further. I don’t want to rain on anyone’s parade but there is another way of looking at these results.
Firstly, only 1 of these new drugs has had any safety studies done; the other 2 could just be too toxic for humans.
Secondly, having screened over 107million compounds they only found 3 new antibiotics! That’s only 1 new antibiotic for every 35 million compounds screened; so yes, we still need to try and protect the antibiotics we have for a bit longer after all.
Thirdly, pharmaceutical companies would still have to take these new compounds through clinical trials in order for them to be used in humans and this will still take years and lots of money. Pharmaceutical companies just aren’t interested in developing new antibiotics; there’s no money in it.
So I’ll complete my CPD folder for this year (at least) as it seems Microbiologists still have a role after all and we can’t just all go home and leave it to the computer geeks.
Now how do I get shares in whoever is making Halicin…?! (Honest I receive no financial incentive to write blogs, I do it because the Editor-Chief-in-Charge finds me news stories to investigate!)
A Deep Learning Approach to Antibiotic Discovery. Stokes J, Yang K, Swanson K et al. Cell Feb 2020, 180: 688-702