Machine learning and artificial intelligence (AI) are already being actively used in drug discovery to evaluate potential binding of small-molecule drugs to proteins, but there's potential for the technologies to be used on the development side as well, especially in hard-to-treat diseases such as Alzheimer's disease.
"I think the drive in this field is that we saw the success of Google and Facebook to basically take what seemed like white noise out in the ether and squeeze out this gold; it's like the new gold rush," Newman Knowlton, a statistician at Millcreek, Utah-based statistical consulting firm Pentara Corp. told BioWorld. "And in Alzheimer's we've been desperate for success, and it seems like we have this messy, noisy data – it almost feels like static sometimes – and we want a machine learning algorithm to come in and wrangle that up and squeeze gold out of what looks like noise. Unfortunately, I don't think it's going to be that easy."
That said, Knowlton sees why people are "jumping on the buzzwords" and incorporating AI and machine learning into development. The addition can generate excitement, increasing funding and the number of participants in clinical trials.
AI could be helpful in diagnosing Alzheimer's disease because it's good at finding patterns that humans miss. Changes in biomarkers or speech patterns might show a propensity towards developing Alzheimer's disease. "It doesn't really matter if the patterns are human interpretable. If they're accurate at predicting Alzheimer's, then I don't care if I can figure out why," Knowlton explained.
But for drug development and clinical trials, patterns that show improvement need to be interpretable by humans. "If my uncle has Alzheimer's disease and I feed it to a machine learning algorithm and it says, oh no, he's getting better, the drug is working, but I can't notice any improvements, and he can't notice any improvements, it's not useful," Knowlton said.
Where Knowlton thinks technology could be helpful in drug development is in rating outcomes of patients. If a machine learning algorithm could be trained to rate patients, it could theoretically be more accurate and reduce the variation that's seen between raters.
Unlearn.AI Inc., of San Francisco, has developed a process to create digital twins of Alzheimer’s disease patients in the active-treatment group that predict how they would perform if they weren't given drug.
The company's Digenesis process uses historical clinical trial datasets from patients who participated in the control arms of clinical trials to create the twins who are a mash-up of multiple patients in the studies.
In a paper published in Nature Scientific Reports earlier this year, Unlearn.AI described using Conditional Restricted Boltzmann Machine to simulate patient outcomes based on 44 clinical variables from 1,909 patients with mild cognitive impairment. The model was then validated using clinical trial results for patients who weren't used to generate the model. "The digital subject data that are generated by the machine learning model are statistically indistinguishable from actual subject data." Charles Fisher, Ph.D., founder and CEO of Unlearn.AI, told BioWorld.
Using multiple patients to create the digital twin makes the process more accurate than a typical matched historical control. Given all the variables, such as age, gender, genetic make-up and performance on cognitive tests, it's impossible to find a historical patient that matches all the characteristics of the patients in a study. Creating a digital twin "removes variability and bias that you typically have when you apply direct historical controls where you end up with some differences between the external control population and the study population," Fisher explained
The digital twins can also be used to generate data beyond what's available in the clinical trial datasets. At the 12th Clinical Trials on Alzheimer’s Disease Meeting in San Diego earlier this month, Unlearn.AI presented an updated model using an improved dataset of about 5,000 Alzheimer's disease patients that also looked at additional factors, such as genetic makeup and concomitant medications.
The model was used to predict results out to 48 months in the TEAM-AD clinical trial even though the model was only trained on 18 months’ worth of data.. "We're able to take all that knowledge that's in those historical datasets and produce simulations that are relevant for new kinds of clinical trial designs or new kinds of patient populations," Fisher said.
Unlearn.AI is currently working on validating its model prospectively in clinical trials and is in active conversations with regulators to convince them that the digital twins should be used in pivotal studies. The company is proposing that the digital twins be used to supplement smaller placebo control arms rather than eliminate them entirely. How much the added machine-learning data can reduce the size of clinical trials, as well as the time frame for a decision, are still up for debate Fisher said.
Unlearn.AI is also working on a publication predicting outcomes for multiple sclerosis patients, so the company is very interested in getting a framework from regulators on the incorporation of the technology in clinical trials. "It's not entirely clear where we fit. We're hoping that we can get some clarity about that from FDA," Fisher said. "I really think there should be some type of overarching regulatory framework to ensure that companies – not just us , but all companies that are trying to use machine learning and AI to improve drug development – that there's some sort of standards applied to all of us to ensure that these tools are working as expected."