Scientists Unveil the First Individual Risk Prediction Model for Multiple Myeloma
New machine learning method could improve predictions for treatment response, prognosis.
For those with a new diagnosis of cancer, the future can be dauntingly murky.
That’s true not just in a philosophical sense, but in a statistical sense as well—most methods of predicting patient outcome are based on probabilities and averages, some of them not very precise.
Many of these prognostic methods are akin to a weather forecasting model, said C. Ola Landgren, M.D., Ph.D., chief of the Division of Myeloma and director of the Sylvester Myeloma Institute at Sylvester Comprehensive Cancer Center at the University of Miami Miller School of Medicine.
“If you want to know, is it going to rain today, you might see in the forecast that there’s a 30% probability of rain,” Dr. Landgren said. But that level of uncertainty isn’t acceptable in all scenarios—take a pregnancy test, for example. “You’re not going to accept a 30% probability of pregnancy. You want to know, am I pregnant or am I not?” he said.
Now, Dr. Landgren and a team of researchers at Sylvester and collaborating institutions have unveiled a new computational model that aims to reduce that uncertainty of the future for newly diagnosed multiple myeloma patients. The model, described today in a publication in the Journal of Clinical Oncology, is the first to predict an individual’s personalized prognosis based on the patient’s tumor genomics and their treatments.
A Precision-medicine Future
The multiple myeloma field desperately needs better prediction tools, Dr. Landgren said, as the number of treatments for the disease has dramatically expanded in the past two decades. That’s great news for patients who are now living much longer with the disease than in decades past. But with so many options, clinicians need better ways to determine which treatment is going to work best for each patient. It’s not possible to just try one therapy after another; physicians need a way to predict which treatment will work or fail before that failure happens.
“The future of the field will have to be focused on precision medicine,” said Dr. Landgren, who is senior author on the new publication. “There’s no other way forward.”
To build the model, the Sylvester researchers and their collaborators used genetic, treatment and clinical data from nearly 2,000 patients newly diagnosed with multiple myeloma. From sequences of the patients’ DNA, the scientists first identified 90 “driver genes”—genes bearing mutations in the cancer cells that appear to spur tumor growth. They then looked at the treatments each patient in their dataset received and how the patients fared on those treatments, matching treatment outcome to an individual’s tumor genetic sequences.
A Model for the Community
The resulting computational model, dubbed the Individual Risk Model for Myeloma or IRMMa, does a better job predicting outcome than previous prognostic tools due to its focus on individual tumor genetics paired with treatment outcomes, the researchers said. Because it’s built using machine learning, it can also learn and improve the more data it receives. The model is built in a manner that allows emerging datasets with future treatment strategies to be added. The research team is now working on including additional datasets from patients treated with newer antibody-based multiple myeloma therapies.
To gather such a large dataset, the researchers needed a large set of collaborators. The Sylvester team worked with scientists at Memorial Sloan Kettering Cancer Center, NYU Langone Health, Moffitt Cancer Center, Heidelberg University Hospital and the Multiple Myeloma Research Foundation to amass the patient data. The model is available online for anyone to use, although its current iteration is aimed at researchers rather than clinicians, said Francesco Maura, M.D., an assistant professor at Sylvester and first author on the study. The model could be useful for researchers interpreting or designing new clinical trials, for example, to provide a large set of comparisons to the experimental treatment. Dr. Maura and his colleagues also aim to improve the model by integrating more patient data.
“This model can only grow with the help of the research community,” Dr. Maura said. “The next challenge is to keep feeding it with the right datasets so at a certain point it will be usable for clinical purposes.”
Tumor Biology over Quantity
IRMMa improves on previous prognostic tools because it takes into account the biology of patients’ tumors, Dr. Landgren said. That’s important in many cancers, but especially so in multiple myeloma, which is highly variable. In fact, their analyses identified 12 distinct subtypes of the disease, a classification which hadn’t been made before.
The original method for classifying multiple myeloma was based on staging for solid tumors and developed in the 1970s. This method relied on the amount of cancer present — more cancer in the body meant a worse prognosis. With the treatments available at the time that staging method was developed, it was fairly accurate, Dr. Landgren said. But with newly developed therapies, especially immunotherapies, the amount of cancer is often less important than the nature of the cancerous cells. Different kinds of driver mutations in the tumor genome affect the cancer’s growth, so certain subtypes of myeloma could have a very good outcome even if they’re diagnosed when the cancer is widespread, assuming the right treatment is matched to the patient.
Even as updated prognostic tools came on the scene since the 1970s, the field was still lacking precise prediction, Dr. Maura said. Some of these tools included tumor genetics, but there have been recent new findings about genetic risk factors that hadn’t yet been included in a prediction model. And all the tools to date relied on averages of a population, lumping patients into groups like “standard risk” and “high risk” and giving prognoses for those groups overall that didn’t take into account individualized risk and how this can be modified by distinct treatments.
“Our model is based on the idea of predicting the risk of the individual patient rather than that of the group,” Dr. Maura said.
IRMMa is also flexible—a patient’s prognosis from the model can be changed if, for example, they receive a transplant after a given treatment. When new therapies become available, as long as there is data from at least a few hundred patients, the model can be updated to incorporate those treatments.
And while the field isn’t quite at the point of sequencing entire tumor genomes for every newly diagnosed patient, that might come in the near future as whole genome sequencing becomes more economical, Dr. Landgren said.
“More and more information will become available, and tools like this model are the future for optimized treatment and management,” he said.
Tags: Dr. C. Ola Landgren, Dr. Francesco Maura, multiple myeloma, Sylvester Comprehensive Cancer Center