Abstract B017: Analysis of miRNA homology using graphs and applications to ovarian cancer prediction

Abstract Introduction: miRNA is a promising biomarker for the prediction and diagnosis of ovarian cancer. We present a novel method to interpret and improve consistency of miRNA cancer prediction models using graph theory. Our technique performs dimensionality reduction independent of relative expression of miRNA species by using only the nucleotide sequences of miRNA, i.e., two miRNA have similar k-homology if they share a common length k nucleotide subsequence. We use this idea to build a graph which describes miRNA similarity, i.e., for some k we let each miRNA be a node of a graph, and the nodes are connected if they have similar k-homology. Methods: The graph theory concept is applied to two large miRNA data sets, one from Japan (n = 6193, p = 2497) and the other from the TCGA database (n = 10549, p = 2497), where p denotes the number of miRNA considered. The TCGA data is drawn from tissue and contains samples from 32 different cancer types and normal tissue. The Japanese data is drawn from blood and comprises 13 different cancers and controls. After performing hierarchical clustering on the graph, we group the miRNA into a smaller collection of families, where each family of miRNA have similar k-homology. We then use this data to train a depth convolutional neural network for cancer prediction. The key difference between the proposed and a standard Feed-Forward Network (FFN) is that we input essentially one feature per family into the dense layers as opposed to inputting all miRNA separately as independent variables, which helps reduce redundancy and improve regularization of the network. Results: On the Japanese data, the proposed model (setting k = 8) offers a mean and standard deviation AUC score of µ = 0.971 and σ = 1.7% compared to µ = 0.961 and σ = 3.6% using a simple 1-layer FFN. On the TCGA dataset, µ = 99.6%, σ = 0.6% using the proposed, and µ = 99.5%, σ = 0.8% with FFN. When applied to ovarian cancer prediction, our method offers AUC = 0.96 and AUC ∼ 1 on the Japanese and TCGA data, respectively, when classifying ovarian cancer samples against all other classes. Conclusions: Beyond improving consistency of neural network models, our technique offers a new means to interpret such models and analyze relations of miRNA to cancer. By assessing the model weights, we can determine the family of miRNA most closely related to a given cancer type of interest. With respect to ovarian cancer, we can identify the nucleotide strings which are most commonly associated with ovarian cancer and analyze how they relate to the 2-D structure of the miRNA. For example, when k = 8, the family of 14 miRNA with similar k-homology to miR-1185-1-3p, has strong relation to ovarian cancer. The k-nucleotide sequence CAGGGGGA is common to 5 of the 14 miRNAs, and, regarding 2-D structure, intersects the bulged base on the upper stem of the pri-miRNA in each case. This tool provides further insight into the mechanics of miRNA in relation to cancer and may improve understanding of miRNA-based neural network models. Citation Format: James W. Webber, Kevin M. Elias. Analysis of miRNA homology using graphs and applications to ovarian cancer prediction [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Advances in Ovarian Cancer Research; 2025 Sep 19-21; Denver, CO. Philadelphia (PA): AACR; Cancer Res 2025;85(18_Suppl):Abstract nr B017.