Abstract A045: Improving the diagnostic accuracy of an ovarian cancer triage test using a joint miRNA-protein model

Abstract

Abstract Background: Multianalyte protein assays that combine several protein biomarkers with an individual’s menopausal status can aid in the preoperative prediction of ovarian cancer among women presenting with an adnexal mass. Serum microRNAs (miRNAs) are another class of biomarkers which have also shown potential to discriminate ovarian cancer cases from benign lesions or healthy controls. Here we investigate the complementarity of miRNA and protein-based approaches. Methods: The population consisted of serum samples from n=678 total study subjects. The training set was provided by Aspira Women’s Health, comprising n=568 women undergoing preoperative evaluation for an adnexal mass (n=333 benign and n=235 malignant). An independent, external validation set (n=110) comprised study subjects enrolled in prospective collection protocols at Brigham and Women’s Hospital, including n=59 subjects with benign masses or healthy controls and n=51 ovarian cancer cases. Among the 51 cases in the validation set, n=20 were FIGO Stage I/II and n=31 were FIGO Stage III/IV. Histologies included n=34 serous and n=17 non-serous. All samples underwent miRNA profiling using a custom 179-miRNA panel optimized for serum analysis using the Fireplex® circulating miRNA assay (Abcam, Cambridge, MA). Proteins were measured using ELISA kits by Aspira. Seven proteins were provided for each patient, including CA-125, as well as age and menopausal status. To reduce the dimensionality of the miRNA data, we employed a forward regression algorithm, which selects a subset of miRNA to optimize linear separation between cases and controls. The miRNA panel was combined with the protein and metadata to train a neural network to diagnose ovarian cancer. We used receiver operating characteristic curves to compare the effectiveness of models trained on miRNA + protein + metadata (joint) vs miRNA alone vs protein + metadata as assessed by the area under the curve (AUC). Results: The joint model produced an AUC 0.95 on the validation set. The miRNA and protein + metadata models offered an AUC of 0.84 and 0.9, respectively. By histology, the joint model offered the highest sensitivity among cancers which were serous and early-stage (89%) and among early-stage cancers overall (90%) when compared to miRNA alone (56% and 80%) or protein+metadata (67% and 65%). Conclusions: These results suggest that miRNA, protein, and metadata are complimentary tools, and the proposed model, which uses all data types simultaneously, is shown to offer the optimal AUC on the external validation set, when compared to the same model trained on miRNA alone or proteins and metadata. The joint model appears to be particularly effective for early-stage, high-risk histologies. This data will be useful to inform ovarian cancer screening and early detection efforts. Citation Format: James W. Webber, Laura Wollborn, Sudhanshu Mishra, Allison Vitonis, Daniel W. Cramer, Ryan Phan, Todd Pappas, Dipanjan Chowdhury, Kevin M. Elias. Improving the diagnostic accuracy of an ovarian cancer triage test using a joint miRNA-protein model [abstract]. In: Proceedings of the AACR Special Conference on Ovarian Cancer; 2023 Oct 5-7; Boston, Massachusetts. Philadelphia (PA): AACR; Cancer Res 2024;84(5 Suppl_2):Abstract nr A045.