Introduction: Previous studies have just suggested that ovarian cancers display unique expression patterns of noncoding RNAs. We hypothesized that circulating microRNA (miRNA) profiles might have utility for the noninvasive diagnosis of ovarian cancer.

Methods: We used miRNA-seq to construct complete miRNA serum transcriptome profiles for 179 women seen at Brigham and Women’s Hospital prior to surgical evaluation of a pelvic mass. Samples were divided 3:1 into a 135-patient training set and 44-patient validation set. Samples included all stages and histologies of ovarian cancer as well as benign pelvic masses and healthy controls. Thirty-three variants of machine learning algorithms were explored to separate the cases of invasive cancer from healthy controls or benign/borderline masses. Tools were graded in terms of receiver operating characteristic area under the curve (ROC AUC). Once the model inputs were established, the specificity of the algorithm for ovarian cancer was confirmed using an independent publicly available dataset of 454 patients with a range of diagnoses, including ovarian cancer. The algorithm was then recalibrated for qPCR using a semi-independent dataset of 346 patients (120 samples overlapping with the original dataset plus 226 additional samples), and externally validated a second time using a completely independent set of 51 samples from Poland.

Results: The optimal machine learning algorithm derived from sequencing was a 14-member miRNA neural network based on miRNA fold changes with an AUC of 0.90 (95% CI: 0.81-0.99). The algorithm performed equally well among early-stage and advanced-stage cases as well in premenopausal and postmenopausal samples. In the independent dataset, the neural network perfectly classified patients in the training set (AUC 1.00, 95%CI 1.00-1.00) and provided very good discriminatory power on the testing set (AUC 0.93, 95%CI 0.81-1.00), with an overall sensitivity of 75% and specificity of 100%. Moreover, the miRNA signature was unique to ovarian cancer compared to 6 other cancer types. qPCR measurement of the miRNA inputs followed by a global sensitivity analysis permitted simplification of the neural network to 7 target miRNAs plus 2 normalizers. On the external validation set, this test had a positive predictive value of 91.3% (95% CI: 73.3% - 97.6%) and negative predictive value of 78.6% (95% CI: 64.2% - 88.2%).

Conclusions: A properly trained neural network derived from miRNA sequencing and tested on more than 900 patient samples can reproducibly discriminate both early- and late-stage invasive ovarian cancer cases from healthy controls, benign pelvic masses, and other types of cancer. This offers the potential for a novel testing strategy for this disease.

This abstract is also being presented as Poster B10.

Citation Format: Kevin M. Elias, Wojciech Fendler, Konrad Stawiksi, Allison F. Vitonis, Ross S. Berkowitz, Gyorgy Frendl, Magdalena Kedzierska, Panagiotis A. Konstantinopoulos, Daniel Cramer, Dipanjan Chowdhury. Derivation and validation of a serum diagnostic test for ovarian cancer using miRNA-seq. [abstract]. In: Proceedings of the AACR Conference: Addressing Critical Questions in Ovarian Cancer Research and Treatment; Oct 1-4, 2017; Pittsburgh, PA. Philadelphia (PA): AACR; Clin Cancer Res 2018;24(15_Suppl):Abstract nr PR09.