Longevity Briefs: How AI Found A Drug Success That Humans Missed

Longevity briefs provides a short summary of novel research in biology, medicine, or biotechnology that caught the attention of our researchers in Oxford, due to its potential to improve our health, wellbeing, and longevity.

The problem:

If used correctly, artificial intelligence (AI) technology has the potential to greatly accelerate scientific research. One of the ways it is already being used is to detect patterns within scientific data that humans missed, as they would have no reason to look for them. For example, AI is being used to examine vast amounts of data from past clinical trials of failed drugs and investigate whether some of the effects of those drugs might be useful in diseases other than those that were being targeted at the time.

One difficulty with clinical trials is that a drug might work in some subcategories of patients but not others. A drug that was only beneficial in 10% of Alzheimer’s patients would still be valuable to that subset of people. However, unless a clinical trial is able to correctly identify such a subset and analyse the effectiveness of the drug for them separately, this treatment would be likely to be deemed ineffective. This is particularly problematic for Alzheimer’s as current methods for predicting the severity and progression of the disease are poor. An AI trained to predict Alzheimer’s progression might be able to re-analyse such data, stratify the participants into subgroups, and detect the effect. That’s exactly what has been done in this study.

The discovery:

Researchers developed a powerful AI-guided tool called a predictive prognostic model (PPM). This model uses machine learning to analyse multiple types of data from patients, including brain scans (MRI), amyloid levels (PET scans), genetic information (APOE4) and cognitive test results. By combining this data together, it was able to accurately predict whether a patient would remain clinically stable in the years following their initial assessment, or whether they would experience a rapid decline in cognitive function. When tested on data where patient outcomes were known, the model achieved a sensitivity of 87.5% (it correctly predicted 87.5% of patients who would go on to experience rapid decline) and a specificity of 94.2% (it correctly predicted 94.2% of patients who would go on to remain stable). For more about what these values mean, why they’re different and why it matters, see our article on the subject.

The researchers then applied their model to data from the AMARANTH clinical trial. This was an unsuccessful trial of a drug called lanabecestat, designed to prevent the buildup of amyloid plaque, in 1354 early Alzheimer’s disease patients. The PPM model was used to classify AMARANTH participants into “slow” and “rapid” progressors based on their characteristics at the start of the trial. The effects of lanabecestat were then re-analyzed in both groups. Remarkably, the slow progressors (representing about a third of the participants) treated with the higher dose of lanabecestat (50mg) showed a 46% slowing of cognitive decline compared to the placebo group, while the rapid progressors did not show a statistically significant benefit from treatment. This suggests that lanabecestat, thought to be an ineffective drug, may actually have been effective in a subgroup of patients that AI-guided models can be trained to identify.

The researchers also re-classified participants into slow and rapid progressors 52 and 104 weeks into the trial to see how many slow progressors would transition into rapid progressors (and therefore become less treatable). They found that in the 50mg treatment group, a majority (nearly 67%) of slow progressors remained slow progressors after 104 weeks. In the placebo group by contrast, only 40% of slow progressors maintained slow disease progression after 104 weeks, with a majority becoming rapid progressors.

The implications:

Real world populations of living humans are diverse and complicated. When it comes to developing new medical treatments, getting a drug to the clinical trial stage is only part of the battle. The way in which data is analysed is very important, and this study shows us how AI tools could help us to improve this analysis and to detect beneficial effects of drugs that only apply to certain people. This could make Alzheimer’s trials (and clinical trials as a whole) more efficient, and therefore more profitable to conduct.

One possible ‘concern’ is that AI might be used to identify those most likely to benefit from a drug, so that said drug can be trialled only on that subgroup as a way of cutting costs. For example, the researchers estimated that if a future Alzheimer’s trial were to recruit only patients classified as slow progressors, the number of participants required to achieve statistical significance would be 90% lower. This could conceivably lead to the reverse situation where data indicating whether a drug works on rapid progressors is delayed. The counterpoint to this is that clinical trials already do this where possible. After all, should we not start by testing a drug on the people it is most likely to benefit so that successes can be delivered as rapidly as possible? There is a reason that clinical trials targeting Alzheimer’s usually look at people in the early stages of the disease – they are the easiest to treat.