Using Machine Learning to Predict Whether Immunotherapy Will Work

For the new favorite in cancer treatment, immunotherapies, arguably the main disadvantage is the difficulty of predicting who will benefit. A research team uses machine learning , an artificial intelligence application, to identify better responders. In a recent study by Leiserson and colleagues in PLOS One in December 2018, the team used data from a clinical trial involving bladder cancer patients and accurately predicted which patients would mount an immune response. The approach was also shown to potentially prevent half of overtreatment.

Using a patient’s own immune system to treat cancer, immunotherapies have produced revolutionary results in recent years, but most work for only a minority of patients. Better use and reduced costs depend on more accurately predicting who benefits.

University of Maryland computer science professor Mark Leiserson and researchers from Microsoft Research and Memorial Sloan Kettering Cancer Center tested, to better predict immunotherapy responses, an approach analyzing extensive data about patients and their tumors through a new computer model.

The multifactorial computer model developed by Dr. Leiserson’s team included % 38 of non-beneficiaries but % 100 of beneficiaries. Thus, accuracy was improved with emphasis on those benefiting. Importantly, it used 3 different types of data not considered in standard cancer treatments and studies: tumor genetic data, immune cell data, and patient clinical data.

Until now, immunotherapy research had focused on finding “one” biomarker to predict who responds. Perhaps this was the mistake, because one or a couple of biomarkers were insufficient. The problem appears solvable by analyzing multiple factors through machine learning.

To build the models, Dr. Leiserson’s team analyzed bladder cancer clinical trial data with an exceptionally rich dataset collecting tumor-cell, immune-cell, and patient information. They defined 36 different features in the models.

On the left of the figure below, 3 different categories containing 36 features can be seen:

The resulting algorithm, with features analyzed together, explained % 79 of patients’ immunotherapy responses using 20 identified features. No biomarker had previously predicted responses at this level.

Researchers also found that removing any of the three data categories (tumor, immune-cell, or patient clinical data) made immune response no longer predictable, with predictive power falling to % 23. Dr. Leiserson emphasized the importance of a multifaceted approach rather than these particular 20 features.

These may not be the only features useful for predicting immunotherapy response. They may change, but the three categories encompassing tumor, immune-cell, and patient clinical data must be considered.

In conclusion, this study addresses tailoring treatments to the genetic and molecular profiles of patients’ tumors, as pursued in precision oncology and makes an important contribution to current efforts in this area.