CHICAGO – Duke University School of Medicine researchers have improved the performance of a model for determining metastatic castrate-resistant prostate cancer (mCRPC) patients' overall survival prognosis by adding circulating-tumor DNA pathogenic genetic alterations to clinical factors.
Presenting at the American Society of Clinical Oncology's annual meeting on Saturday, Susan Halabi, a professor of biostatistics and bioinformatics at Duke, reported that the researchers built on a previously developed model that used clinical factors to determine overall survival in patients with mCRPC who had been treated with androgen receptor inhibitors. The model originally factored in performance status, disease site, opioid analgesic use, and the presence of lactate dehydrogenase, albumin, hemoglobin, prostate specific antigen, and alkaline phosphatase, Halabi noted.
Researchers sought to update this model with pathogenic genetic alterations identified in ctDNA in 776 mCRPC patients' samples from the previously conducted Alliance A031201 Phase III trial. This study tested the activity of Pfizer and Astellas' Xtandi (enzalutamide) with or without Xtandi and Janssen Biotech's Zytiga (abiraterone) and prednisone.
The original model using only clinical factors had an area under the receiver operating characteristic curve of 0.72, while the model with combined clinical and genetic factors had an area under the receiver operating characteristic curve of 0.77 — a "statistically significant" increase in accuracy, Halabi added.
To determine which genetic factors would be included in the updated model, Halabi said the team used the 69-gene AR ctDetect assay, which utilized cell-free DNA isolated from up to 3 milliliters of plasma. The University of Minnesota Genomics Center performed DNA sequencing library preparation and custom targeted panel sequencing, and the researchers developed a custom bioinformatics pipeline to identify ctDNA aneuploidy fraction, copy number gains or losses of target panel genes, pathogenic single nucleotide variants in target panel genes, and AR gene structural rearrangements.
In building the model, Halabi's team used a machine learning-based approach to first identify the genetic variables that were associated with survival in mCRPC patients. They cross-validated the model by randomly splitting the data 100 times to check its performance and found that the variables "were very much robust," Halabi said.
The researchers incorporated the clinical variables, ctDNA aneuploidy fraction, and BRCA2 loss into the model and determined the time-dependent area under the receiver operating characteristic curve to assess the model's predictive accuracy, Halabi added.
In the updated model, the genetic features that were included were AR enhancer gain, hemoglobin, MYC gain, RSPO2 gain, and alkaline phosphatase, she noted.
Finally, the researchers used the predictive risk score generated by the updated model to categorize patients in the Alliance A031201 trial into prognostic risk groups. For example, when researchers use the model to stratify patients into four prognostic risk groups — low, low intermediate, intermediate poor, and poor — those in the poor risk group had a median overall survival of 16.3 months while some patients in the low risk group were still alive when the trial ended. When they stratified patients into a three-prognostic risk group (low, intermediate, and poor), the poor risk group had a median overall survival of 17.9 months, while some in the low risk group were still alive.
The model "could be used not only for counseling patients, but it could be used in the design of randomized studies for selecting patients for an enriched design or as a stratification variable … in the randomization," Halabi said.
She highlighted that this genetic analysis was possible using a small amount of plasma garnered from noninvasive testing but added that the model needs to be externally validated, particularly "in the context of other therapies."