Despite costing more than US$1 billion each, and taking up to 15 years to develop, 90% of new human drugs under development fail in clinical trials.
But leading CSIRO experts are using artificial intelligence (AI) in the search for new antibiotics and say it will lead to a better hit rate of successful pharmaceutical candidates.
“A lot of the challenge now is in actually picking the winners,” says Dr Lewis Blackman, Research Team Leader of CSIRO’s Drug Discovery Chemistry Team. “In some cases, you can get something more rapidly into pre-clinical trials and, in the future, maybe even into clinical trials, but a lot of the cost is in that clinical-trial space, so how do you select good winners? I think AI is where the field is moving.”
Dr Blackman says one of the chief advantages of AI is its ability to search through millions of datapoints to find, for example, how different compounds inhibit the growth of E. coli, and then rapidly make predictions about the effect of other compounds. “The neural network will try to align the given inputs and outputs,” he says. “And in this case, the inputs are the chemical features and the outputs are the biological activities. So, it will ask, ‘Does the compound inhibit bacterial growth, yes or no?’ And during training, it will change the weightings of all of these equations, all of these nodes in the network, to try to get the best alignment.”
“
Future models might be able to filter out compounds that may look good in vitro, but aren’t going to succeed in an in vivo setting.
Lewis Blackman
The advantage, says Dr Blackman, is the process finds different patterns that researchers aren’t ordinarily able to discover on their own just by looking at the data set.
“Future models might be able to filter out compounds that may look good in vitro, but aren’t going to succeed in an in vivo setting,” he says.
Dr Denis Bauer, Group Leader of CSIRO’s Transformational Bioinformatics Group, develops AI techniques to research everything from virulent COVID strains to the evolutionary trajectory of the flu. She says AI is a great tool, but not necessarily the cure-all when it comes to AMR. “Everybody thinks it’s going to be the silver bullet, but machine learning and the precursors of AI have been around for 30, 40, 50 years. So, it’s not going to be the new thing that is going to revolutionise everything,” she says. “But what has changed is the capability and the power of the computers that are actually running it. So, we can train larger, more involved models.
“The other thing that has changed is data. There’s more data, there’s more actual evidence – observations from the real world that we can train our methods on.”
Constantly changing targets
“
There are million ways in which nature can overcome this threat to the bacteria.
Denis Bauer
One of the special challenges of AMR research is that the bacteria can evolve faster, for example through horizontal gene transfer. “If you sequence, let’s say 5 million E. colis, that does not give you the same level of depth and understanding of the full breadth of evolutionary scope as it would give you for 5 million humans,” Dr Bauer says. “Because with E. coli the 5 million sequences today might be totally outdated come next year, after they have had interactions with other bugs.”
In addition, as more is understood about AMR, it is clear there isn’t one mechanism that makes bacteria resistant to a specific drug. “There are million ways in which nature can overcome this threat to the bacteria.”
Dr Bauer says another challenge with AMR is that, historically, data about bacteria has been stored in terms of their ‘clades’, or family groupings. But that sort of classification may not be ideal when trying to find mutations that confer drug-resistance or that make them more pathogenic. “It’s too crude of a method to put them in drawers because of the horizontal gene transfers, rearrangements, and all the other weird and wonderful things that these bugs do with their genomes in order to test the full breadth of the evolutionary space,” she says.
Data across borders
Critical to this AI-driven research is the ability to access data across borders, and Dr Bauer says there are multiple issues preventing that free exchange of information, including a fear of being internationally vilified if an outbreak is announced. “Just because an outbreak was first detected in one country, for example, it does not mean that it has originated from there, let alone stay there,” she says. “In fact, it probably will travel around the world quite quickly and therefore countries need to be able to share the knowledge of those things, rather than be afraid of what the consequences are.”
Commercial constraints can compound the issue further, as data assets may not be able to be freely shared. “These repercussions and commercial disadvantages need to be balanced with massive incentives for information exchange to happen appropriately,” Dr Bauer says.
Dr Blackman says CSIRO is continuing to critically consider the ongoing use of AI in developing strategies against AMR. “I think with a lot of these AI tools the proof will be in the pudding,” he says. “Does it just work in this particular use case? Does it work in others? Can it work with datasets that are maybe noisy, messy, where you haven’t got standardised testing, for example?”
However, he says the use of AI is increasingly considered a typical tool in the drug-discovery development toolkit, useful in certain cases. “The average patient who’s taking these medicines in 20 years’ time is probably not going to know if it’s an AI-derived drug or not. And for all intents and purposes, it doesn’t matter because both AI-derived and conventionally discovered drugs would have gone through exactly the same robust safety and efficacy testing to gain market approval.”
Ken Eastwood is a highly experienced award-winning editor, journalist, author and communicator, with particular expertise in science, agriculture, sustainability and rural affairs.
Other stories in our AI and AMR series
