How This Scientist Used AI To Find A Drug Cocktail Effective Against The Covid-19 Delta Variant
Dean Ho, director of the National University of Singapore's Institute for Digital Medicine, used AI and testing on the live virus to derive an optimal combination of drugs and dosages to treat Covid-19
A team of Singapore researchers have identified a new combination of drugs that can be used to treat Covid-19 patients, showing promising results in its efficacy against both the Beta and Delta variants of Covid-19.
The team, led by Dean Ho, the director of Institute for Digital Medicine—a health optimisation institute within the National University of Singapore—used an artificial intelligence platform called IDentif.AI to identify the optimal therapy.
The platform was set up and first put to use in April 2020, when it identified a combination of three drugs to treat the Sars-CoV-2 virus: lopinavir and ritonavir, which are used to treat HIV patients, and remdesivir, an antiviral drug originally developed to treat ebola.
The drugs were effective, but were not readily available or easily administered—remdesivir, for example, can only be taken through an IV drip. The team’s most recent experiments focused on readily available drugs that could be taken orally.
After testing a range of 12 antivirals and cancer medications, one of the resulting combinations was antiviral drug molnupiravir together with baricitinib, an anti-inflammatory drug.
Ho says that while there isn’t yet data from clinical trials to prove that the drug combination is effective in all phases of the disease, the combination inhibits the virus in laboratory tests, making it suitable for further clinical evaluation. Molnupiravir, which is being trialed by Merck and Ridgeback Biotherapeutics, recently reported promising results from human trials.
We talk to Ho about the implications of the discovery, and how he used AI to find optimised regimens out of millions of possible drug combinations with precision—all in a three-week timeframe.
Can you describe in simple terms the breakthrough you’ve made?
The power of IDentif.AI is that it’s not a prediction approach; it’s an optimisation approach. From millions of possible [drug] combinations, our platform finds the best possible combos. And the cool part is it’s not one combination; we rank them from the best all the way to the worst, which means we give clinicians options for their patients. If one drug is not available, we give them rankable options on what to try next for the patient. What was different about this study is we used new sets of therapies, recommended by our local and amazing clinicians here, that are in tablet form, which means these will be more easily administered to patients in the community, as we prepare for endemic disease.
How did you do this?
We run experiments on the live virus to experimentally determine what the best possible [drug] combinations are. We don’t only look at one drug, or one combination, which is what traditional approaches do. We take large pools of drugs, ten drugs or more, and that represents hundreds of thousands or even millions of possible combinations. What’s neat about IDentif.AI is that with a small number of experiments, maybe a couple of hundred experiments, which, for context, you can get done in a few days, we can determine the very best possible combination. It’s a true optimisation strategy. And from there, what’s neat is even though this is done on a live virus on a plate, the answers that we have gotten from IDentif.AI, we’ve shown to be efficacious in humans.
Can you talk us through the process?
It [starts with] a group of frontline clinicians, who have run some of the most important clinical trials for the drugs that we know of to date. They recommend the initial drugs we should think about, based on their clinical knowledge as well as their expertise in how different therapies work. They provide us with this pool of candidate drugs, and we work with a team of virologists and people who run Biosafety Level Three labs, who can handle the live virus, so we can design experiments on the live virus. It’s almost like experimentally crowdsourcing the live virus to tell us what the truly optimal combinations are, all the way down to non-optimal. We come up with these answers in a really fast turnaround. The whole process is a couple of weeks.
What are the potential implications of this discovery?
IDentif.AI, we’re hoping, will be a supporting platform for the central strategy, the primary strategy of vaccination, detection and tracing. Our aim is to provide another component to the arsenal of possible ways to address endemic Covid-19. When we talk about pandemic preparedness, IDentif.AI is agnostic. It can be used for different bacteria, different viruses. The platform can be rapidly adapted to meet the demands of the next pathogen or the next bug that's going to cause serious medical issues. From this study, we very quickly pinpointed strategies to address Delta. If new variants pop up, we'll be ready again to keep trying new therapies to keep the arsenal current.
How important was AI to the discovery?
Contrary to traditional AI, we don't use big data—and there's a reason for that. The data that we need doesn't exist. If you take 10, 12, 15 drugs , there’s so many different permutations that in order to solve that problem, we need data to be curated and acquired prospectively, meaning we have to actually run the experiments to get the data. It's not data that we have from previous studies, because there may not be any studies that have looked at [these drug combinations]. So we actually need very small amounts of data. And that’s why the answers from IDentif.AI are clinically actionable—because we're using actual experiments to find those answers; it's not theoretical.
All this information is open-source, right?
We've developed a publicly available database, which allows clinicians and drug developers from around the world to look into the recommendations we've come up with, with all the different drug sets that we've evaluated on the live virus.
So if a clinician from Taiwan, for example, needs to start treating more patients, he can go into this database and he can say, “Oh, look, these drugs I have—what's the best possible combo I can get from these drugs?” And if another patient comes in who has liver dysfunction and there are certain drugs they cannot take, then the clinician can punch in other drugs that we've tried that are okay to give this patient, and come up with other options. That's the benefit of a platform like IDentif.AI: we don't go in and just try one combination and hope that it works. We're talking about interrogating this massive galaxy of possible combinations and reconciling all of that into a list that doctors can consult.
And they've used it. We've had a partner in Taiwan initiate and get approval for two different clinical studies off of our database. Delta is really starting to crowd out all the other variants, so we're happy that people are finding our platform to be actionable and it’s helping to guide both clinical care as well as how we develop new drugs.
What are the wider lessons you’ve learned from developing the IDentif.AI platform?
IDentif.AI and its other permutations have been widely validated in human studies for cancer, infectious diseases—such as HIV and TB—and preventing transplant rejection. I think there has to be a total rethink about how we develop medicine. Right now, we spend billions of dollars for a 95 percent failure rate. We're spending a lot of money and not delivering the right therapies to patients. And even when we do, the therapies are not optimised. We could always be doing better.
Answers edited for brevity and clarity.