Researchers at the University of North Carolina and North Carolina State University have developed a computational framework to generate evidence-based optimal cancer treatment decisions informed by a patient’s genomic biomarkers. The findings, which may aid in the development of precision cancer treatments, are published in the Journal of the American Statistical Association.
Naim U. Rashid, PhD, an assistant professor of biostatistics at UNC Gillings School of Global Public Health and UNC Lineberger Comprehensive Cancer Center and the study’s first author, said the goal of the research was to develop and train new machine-learning methods to predict optimal treatment based on big data from large scale preclinical screens in patient-derived xenografts, or PDXs.
UNC Lineberger’s Michael Kosorok, PhD, the W.R. Kenan, Jr. Distinguished Professor of Biostatistics and professor of statistics and operations at UNC Gillings, is the paper’s corresponding author.
Created by implanting part of a patient’s tumor into immuno-compromised mice, a PDX line produces multiple models of the same tumor. This makes it possible for researchers to more efficiently test and evaluate how an individual patient’s tumor responds to different drugs. Molecular biomarkers may be collected on each tumor as well, and can be correlated with treatment response. Data derived from such studies are used to estimate the potentially most effective therapy for a patient.
In this new study, Rashid and his colleagues analyzed data from a large PDX screen spanning five cancers, 1000 PDX lines and 38 unique treatments evaluated.
They used a new machine learning method that was tailored to address several unique aspects of PDX data, such as evaluating responses pertaining to a large number of treatments applied to the same tumor, and to search for predictive biomarkers from a large set of genomic features in this framework. This allowed them to more precisely recommend the best treatment given a set of patient biomarkers. This approach involves tree-based extensions of prior methods for estimating optimal individualized treatment rules, such as outcome weighted learning and a reinforcement learning method called Q-learning.
“PDX studies represent an untapped resource to exploit for estimating optimal individualized treatment rules, which can be used to recommend best potential therapy in new patients,” Rashid said. “This new machine learning method was tailored to address several unique aspects of PDX data, such as evaluating responses pertaining to a large number of treatments applied to the same tumor, and to search for predictive biomarkers from a large set of genomic features in this framework.”
The researchers discovered their novel approach outperformed existing machine learning methods that do not leverage the unique structure of PDX data. “We also learned that that the application of deep learning methods such as autoencoders are helpful for distilling relevant information from a large number of biomarkers into a smaller, more salient number of features while also retaining a similar amount of information,” Rashid said. “This work is important because it provides us a computational framework to formalize and learn evidence-based optimal treatment decisions given a set of patient biomarkers.”
Looking ahead, Rashid said the next steps include investigating whether these results can be validated in a clinical trial and in PDX studies that are ongoing at UNC Lineberger.
Authors
In addition to Rashid and Kosorok, the paper’s other authors are Daniel J. Luckett, PhD, Jingxiang Chen, PhD, Michael T. Lawson, PhD, Donglin Zeng, PhD, UNC Gillings; Longshaokan Wang, PhD, Yunshu Zhang, PhD, and Eric B. Laber, PhD, NCSU; Yufeng Liu, PhD, UNC Gillings and UNC School of Medicine; and Jen Jen Yeh, MD, UNC Lineberger and UNC School of Medicine.