Alternative splicing enhanced with deep machine learning: a novel method to detect ovarian cancer

Abstract

<h3>Objectives:</h3> There is a continued effort to diagnose ovarian cancer in early stages. Alternative splicing analysis may provide a unique signature to discriminate between malignant and benign tissues. RNA-sequencing has enabled genome-wide assessment of alternative splicing. Furthermore, deep-learning prediction models of alternative splicing based on exon-specific sequences features enhances the value of this analysis. The aim of this study is to create a model that would discriminate high grade serous ovarian cancer (HGSC) from normal tube using deep-learning (or artificial intelligence, AI) and alternative splicing analysis. <h3>Methods:</h3> This is a case-control study of patients with confirmed HGSC and patients without personal history of HGSC undergoing salpingectomy for benign conditions. RNA-sequencing was performed on all tissue samples. Resulting RNA sequences were introduced in DARTS (Deep-learning Augmented RNA-Seq) software suite. DARTS created a model of differential alternative splicing aimed to discriminate between HGSC and normal fallopian tube. Performance of prediction models were measured by area under the curve (AUC). <h3>Results:</h3> 112 HGSC and 12 benign samples were successfully sequenced. Prior to deep machine-learning, we found 1,726 differentially spliced single exon transcripts between HGSC and benign samples (p<0.005). These transcripts discriminated benign from malignant groups with an AUC of 88% (95% CI: 0.77-1.0). After deep machine-learning, the performance improved to 91% (95% CI: 78-1.0). 95% CIs of both overlapped. <h3>Conclusions:</h3> Individual exon expression used to assess alternative splicing was analyzed with AI methods and created a model that identified HGSC with an AUC of 91%. This method could be applied to single-cell RNA-seq technologies to create diagnostic tools for detecting early HGSC.