As the use of artificial intelligence becomes more a part of our everyday lives, scientist Jineta Banerjee said that for people living with rare diseases, machine learning can boost access to therapies that can help them live more productive lives, while also supporting caregivers who fight to get their loved ones the care they need.

In her role as a principal scientist with Sage Bionetworks, Banerjee helped journalists in NPF’s Covering Rare Disease fellowship program understand the potential of AI, such as identifying biomarkers to predict whether a patient has a rare disease and sifting through large amounts of data to provide patients with improved therapy targets.

But, Banerjee said, “applying machine learning in rare disease faces many challenges.”

Chat-GPT and other popular AI tools work because of “large language models” – likewise, in a medical context, machine learning requires lots of data in order to find reliable patterns.

“Rare disease, by definition, has few patients or few samples for a specific disease,” she said.

Researchers can combine datasets or augment them with prior knowledge but these methods have not always yielded accurate or stable results.

One way to address the instability is “to artificially increase the exposure of the model to rare samples … with a technique called ‘bootstrapping.’ … [This] makes a big dataset out of the samples that were present in the small dataset by sampling over and over again.

“Another way that we can help machine learning models to be more stable and work better when given an unseen sample is with a technique called transfer learning. In this technique, what happens is we take a large dataset and apply a model to it to understand or to extract patterns or representations from that large dataset, and then use these learnings or these learned representations and apply it to a smaller, much smaller rare disease dataset. And this can help identify correlations between the learned representation and the clinical outcome or the clinical label that is present in this rare disease dataset.”

Because machine learning relies so heavily on access to data, Banerjee says researchers, clinicians and advocates must work together to ensure its ability to improve outcomes for people living with rare disease and their families. That requires a painstaking focus on information gathering.

“There are inherent biases in the data where unrepresented populations are not well-served by the data,” Banerjee said, but even so, she believes “we are at a place where machine learning and artificial intelligence methods are improving by leaps and bounds.”

She said AI can also help caregivers. For instance, a parent of a child with a rare disease could run an academic or research article about the disease through Chat GPT and ask for it to summarize it at an eighth-grade or fourth-grade level and it will do so.

Still, Banerjee encouraged journalists to bring skepticism to AI.

“Whenever we come across biomedical articles that claim that they have applied machine learning and AI models to rare disease, we need to be very vigilant,” Banerjee said. “Were the datasets constructed appropriately? … Were multiple approaches taken to build appropriate models? And were the methods statistically rigorous?”

She also noted that “it will still be some time before we actually see the real, tangible effects of AI and machine learning in making these diagnostics and treatments available in the lower-income countries.”

What’s next?

Banerjee is encouraged by the growing popularity of “multimodal data” – data pulled not simply from clinical notes, but also lab results, DNA sequencing, imaging data and more.

This means that each patient can have a more rich dataset.

“We’ll get much more complete pictures of what a rare disease means for a patient in the near future.”

“The next thing that I am excited about is digital twin technology,” which is created when a computational model is built off a real-life entity. “Digital twin technology is actually quite popular in the engineering world. It is already used in the airplane industry, automobile industry, where these digital twins are used to help test technological advances before those advances are released to the general public. When we think about the digital twin technology in rare disease, I want to emphasize that we are really far away from making a digital twin off a patient … but what we can start thinking about is making these computational models of much simpler entities, like patient-derived cell lines or tissue samples taken from patients.”

Scientists can then model the disease and, for instance, “computationally test how a certain drug would affect a patient tissue sample before applying that drug to the patient themselves.”

In the future, Banerjee said she hopes to see more headlines about early rare disease diagnoses, new treatments and even cures, but to get there “we need a lot of molecular, pharmaceutical, and clinical research to really understand many of these diseases and the underlying cellular signaling” that causes them. “but these aspects of research are very laborious and very time-consuming.” That’s why machine learning is so important.

“Any methodology that is developed for helping the rare disease community can also be helpful for personalized medicine in the long run.”

Access the full transcript here.

This training was sponsored by Fondation Ipsen. NPF is solely responsible for the content.