Machine learning puts precision into medicine


Jimmie Ye, PhD
Jimmie Ye, PhD

The latest investigative approaches can look at individual genes within single cells, multiplied by thousands of cells for each of hundreds of healthy and diseased subjects. These new hypothesis-generating studies can reveal novel and surprising truths about physiology and pathophysiology.

Take lupus, a disease with a complex etiology that likely involves many cell types. The latest findings from interrogating about one million cells from 150 lupus patients is starting to create a clearer view of the cell types that contribute to disease.

“Some of the most promising new treatments for lupus have been targeted towards B cells,” said Jimmie Ye, PhD, Assistant Professor of Medicine at the University of California, San Francisco. “We do not see much difference in patients versus healthy controls in their B cells. What we saw were quite a few differences in their T cells. That is an interesting new avenue for potential therapy for lupus.”

Dr. Ye will describe the state of the art research in massively multiplexed investigative approaches that are hypothesis-neutral during Biomarkers, Precision Medicine & Machine Learning from 9:00 – 10:00 am Tuesday in the Thomas Murphy Ballroom 1 & 2, Building B in the Georgia World Congress Center. These approaches generate huge data sets, about 5,000 times more data per investigation than rheumatology researchers have ever had to work with in the past. And while it can be difficult to work with these enormous data sets, the results can change the course of research and
clinical medicine.

Another salient feature of lupus is the interferion-1 (INF-1) signature, a biomarker that has been recognized for decades. Trial after trial has attempted to modulate different components of the INF-1 pathway in lupus with mixed results.

“We were able to figure out that monocytes are the major producers of this INF-1 signature,” Dr. Ye said. “As far as we know, this is the first definitive demonstration that the lupus signature we have known for so many years is driven by a particular cell type. So you can imagine the next generation of drugs targeting not the broad signature, but the specific cell type that is creating the signature. That level of detail lets you put a lot more precision into precision medicine.”

Katherine Liao, MD, MPH
Katherine Liao, MD, MPH

Katherine Liao, MD, MPH, Clinical Investigator in Rheumatology, Immunology, and Allergy and Associate Professor of Medicine at Harvard Medical School, will take a different approach. Genetic information is vital, but it is also more powerful when you know the phenotype that a particular genotype generates.

“Decades ago, genetic research was less complicated when you were looking at familial diseases,” she said. “You knew the phenotype for this incredibly well-characterized disease the patients all shared. Now, you can look at millions of genetic variants across millions of patients, and how do you know which patients have which conditions? Who has what disease and at what stage?”

Focusing on phenotype led her to electronic health records (EHRs), which are treasure troves of data, all of it unstructured, none of it intended for research.

The key is a developing a set of algorithms that can classify patients with different phenotypes in the EHRs and extract the data in usable forms across multiple institutions. Provider notes are the most useful source, she continued. Most clinicians have been taught to observe and record clinical data in very similar ways.

“You code in a particular way because that’s the way you may have been taught at your institution,” she said. “But in your notes, you are saying what you think the patient actually has and why. We can use natural language processing to turn all that text into useful data. And then we turn to machine learning to look for patterns or features combining all the EHR data whether from codes or notes to predict whether a patient has a condition. We do this now for research.”

Machine learning is already a reality in imaging, Dr. Liao noted, and is making inroads in many other specialties, including rheumatology.

“In the future, whether we want it or not, machine learning is going to become part of our everyday clinical practice,” Dr. Lio said.