Universal Time: 20:02  |  Local Time: 20:02 (0h GMT)
Select your timezone:

Clinical classifiers, trials and the patient voice

Thursday May 13, 2021 - 20:20 to 21:00

Room: Breakout

P-4.1 A clinical ovarian cancer histotype classifier using gene expression data for improved diagnostic accuracy

Lauren Tindale, Canada

Postdoctoral Fellow
University of British Columbia

Abstract

A clinical ovarian cancer histotype classifier using gene expression data for improved diagnostic accuracy

Lauren Tindale1, Derek Chui2, Martin Koebel3, Susan Ramus4, Marina Pavanello5, Samuel Leung2, Cathy Tang1, Tayyebeh Mehrane Nazeran2, David Huntsman1,2,6, Blake Gilks2,6, Holly Harris7, Jen Doherty8, Joellen M Schildkraut9, Stefan Kommos10, Michael Anglesio1,2,6, Aline Talhouk1,2.

1Obstetrics and Gynaecology, University of British Columbia, Vancouver, BC, Canada; 2British Columbia’s Gynecological Cancer Research Program (OVCARE), BC Cancer, Vancouver, BC, Canada; 3Department of Pathology and Laboratory Medicine, Foothills Medical Center, University of Calgary, Calgary, AB, Canada; 4Women's and Children's Health, Faculty of Medicine, University of NSW Sydney, Sydney, Australia; 5The Westmead Institute for Medical Research, Westmead, Australia; 6Department of Pathology and Laboratory Medicine, University of British Columbia, Vancouver, BC, Canada; 7Program in Epidemiology, Division of Public Health Sciences, Fred Hutchinson Cancer Research Center, Seattle, WA, United States; 8Department of Population Health Sciences, Huntsman Cancer Institute, University of Utah, Salt Lake City, UT, United States; 9Department of Epidemiology, Rollins School of Public Health, Emory University, Atlanta, GA, United States; 10Department of Women's Health, Tuebingen University Hospital, Tuebingen, Germany

Background: Ovarian cancer histotypes differ in etiology, molecular profile, treatment response, and clinical outcome, making it imperative to correctly classify patient tumours for personalized treatment. Pathologist classification has vastly improved with the implementation of objective biomarkers and additional improvement may be realized by utilizing gene expression to distiguish histotypes.

Methods: NanoString gene expression data was used to develop a histotype classifier from clinical FFPE tumor sample RNA. Models were trained on 72 ovarian cancer associated genes in 1547 samples. Multiple algorithms and subsampling methods were tested to overcome class imbalance bias due to a dominance of HGSC. Models were compared to gold standard classification of expert gynepathologist review and IHC, were based on 500 bootstrap samples, and evaluated overall and per-class, using accuracy, kappa, and F1-score.

Results: Random forest with smote subsampling, where minority classes were synthetically up-sampled and the majority HGSC class was down-sampled, showed the most accurate classification. Smote random forest kappa=0.83 was close to the interobserver agreement observed among gynepathologists (kappa=0.89) and higher than reported agreement among general pathologists (kappa=0.54-0.67). An additional 427 genes of interest were assayed in subsets of the data, but sensitivity analysis suggested no improvement over the initial 72 genes. Future work to further reduce this gene list and validate the model in two external datasets (n=644 and n=1061) is currently underway.

Conclusions: NanoString-based histotype classification was able to reach a high classification accuracy similar to pathologists with gyne subspecialty training. This model provides an objective and reproducibly option for histotyping in diagnostically challenging cases or sub-optimal biopsy sampling  that can be integrated into existing routine pathology operations.

© 2022 Virtual CCOCR 2021