Personalised medicine has transformed cancer treatment. Chemotherapy used to be little more than carpet bombing with cytotoxic chemicals in the hope that the cancerous cells would be wiped out before all the healthy ones were gone too.
Then came the discovery of the thousands of genetic mutations that each cause their own version of cellular havoc in different organs and cells. Now, using targeted, precise cancer therapies tailored to the individual patient, the chance of success is much, much greater, and the collateral damage greatly reduced.
That personalised treatment approach, aided by technologies such as machine learning and deep neural networks, is now being explored in infectious disease – particularly bacterial infections – to not only improve the outcomes of treatment but also reduce the risk of antimicrobial resistance emerging and spreading.
“When it comes to infectious diseases, I think the assumption is that most of it is down to the pathogen, not the person,” says molecular biologist Dr Mathew Stracy, whose lab at the Dunn School of Pathology at the University of Oxford is focused on exploring the mechanisms of antimicrobial resistance. But there’s growing evidence that there are not only pathogen-specific factors relevant to treatment, but host-specific ones and host-pathogen-specific interactions are as well.
“What pathogens you might be colonised with, and what they’re resistant to, alter your risk of getting a resistant infection,” Dr Stracy says. “Which antibiotics you take, together with which pathogens you’re colonised with, affects what happens afterwards as well.”
First, the pathogen itself. If someone is suspected of having a bacterial infection, they are either treated empirically with a broad-action antibiotic – which tends to happen with relatively minor conditions like urinary tract infections or ear infections – or a sample, such as urine or blood, is sent for culturing to identify which bacterial species is responsible and whether they’re susceptible or resistant to antibiotics.
“
We measure thousands to millions of individual cells in each sample, and what we’re looking for are the complex patterns that arise as a result of the drug having an effect on the cells.
Dr Kieran Mulroney
In the case of bacterial infections of the blood – also known as sepsis – this step is not only critical but urgent, says Dr Kieran Mulroney, a microbiologist and CEO of Australian med-tech company Cytophenix. “There’s between a 7% and 9% risk of death every hour that you don’t have the right antibiotic,” he says. But the process of determining the type and susceptibility of bacteria can be laborious, which is why Dr Mulroney and colleagues are bringing the power of deep neural networks to help.
“We measure thousands to millions of individual cells in each sample, and what we’re looking for are the complex patterns that arise as a result of the drug having an effect on the cells,” Dr Mulroney says. A trained and experienced human can do this, but it takes time and those skill sets aren’t widely available, especially in rural and regional areas.
“
This is where the AI for us was the game changer, because we could now analyse in minutes what would previously have taken hours.
Dr Christine Carson
So instead, they’ve developed a deep neural network that can do that job as well as a trained human. “What we’ve been trying to do is create something that can have that expert capability, that non-experts can use to give them the same power that all of that experience brings,” Dr Mulroney says. “So you feed the data in, and it gives you back simple, actionable results.” And it can do it much faster than a human, says microbiologist Dr Christine Carson, Chief Scientific Office at Cytophenix. “This is where the AI for us was the game changer, because we could now analyse in minutes what would have taken hours,” she says. The company is working towards FDA approval for the technology.
Another approach to detecting potential antimicrobial resistance in a patient is to look at their past; or example, if an individual has had recurrent bacterial infections, what antibiotics they have previously been treated with, and how successful that treatment was. There are now several studies using machine learning and deep neural networks, trained on existing electronic medical record databases to learn the patterns that predict a high-risk of antimicrobial resistance in an individual patient.
Dr Stracy is co-author of one study, which combined whole-genome sequencing of more than 1000 bacterial samples from patients – taken before and after antimicrobial therapy – with a machine-learning analysis of more than 140,000 urinary tract infections and more than 7300 wound infections. They found that applying machine learning to this data set could help predict the likelihood of resistance emerging during a new bout of infection – specifically urinary tract infection – even if the patient was treated with antibiotics that their current infection was susceptible to.
This approach could also be used for more serious infections, such as sepsis, Dr Stracy says. “In hospitals, they will typically culture suspected bloodstream infections, so there is more data on those, so that’s another one where you do get these data-driven approaches,” he say
“
It is the really big goal in all of this: how do we tailor our dosing regimens to be optimised to suppress resistance.
Dr Os Cotta
Personalising antibiotic therapy also means tailoring treatment to the patient’s physiology, which again is something not routinely done with most treatment approaches. That’s particularly relevant for patients in intensive care, whose ability to metabolise drugs can be severely altered by their illness, says Dr Os Cotta, a clinical pharmacist with the University of Queensland Centre for Clinical Research. “It might fluctuate to the point where, because their kidneys are shutting down, which can happen in sepsis, they may go from having sub-therapeutic or lower concentrations than are desirable, to much higher concentrations that can put them at risk of having toxicity of the drug,” Dr Cotta says.
Both these scenarios can not only lead to worse outcomes for the patient, but can also create the environment for resistance to emerge and spread to other patients. This is where AI technologies, such as machine learning can once again help, by analysing large amounts of data, looking for patterns that can predict which patients are at a higher risk of antibiotic treatment failure. “Having that ability of machine-learning based methodologies can provide a way of perhaps more rapidly or more efficiently identifying those at risk of antibiotic treatment failure,” Dr Cotta says. “It is the really big goal in all of this: how do we tailor our dosing regimens to be optimised to suppress resistance.”
Bianca Nogrady is an award-winning science journalist whose reporting on science, health and the environment.
Other stories in our AI and AMR series
Accelerating AI to search for new antibiotics
