Charles M. Nichols

Utilizing Serum-Derived Lipidomics with Protein Biomarkers and Machine Learning for Early Detection of Ovarian Cancer in the Symptomatic Population

Abstract Ovarian cancer is the fifth leading cause of cancer-related deaths among women. Most patients are diagnosed at late stage (III/IV), resulting in a 5-year survival rate below 30%. This is driven by the presentation of vague abdominal symptoms that confound diagnosis at early stages (I/II) and a shortage of robust biomarkers. We are taking a novel approach for earlier ovarian cancer detection, leveraging lipids as biomarkers. We utilized untargeted ultrahigh pressure liquid chromatography–mass spectrometry to analyze sera from two large, independent cohorts (N = 433 and N = 399) designed to reflect the symptomatic population, including individuals with benign adnexal masses, early- and late-stage ovarian cancer, gastrointestinal disorders, and otherwise healthy women seeking care for symptoms. We identified a significantly altered lipid profile in ovarian cancer and early-stage ovarian cancer specifically across both cohorts compared with controls. We also profiled select protein biomarkers (cancer antigen 125, human epididymis protein 4, β-2 folate receptor α, and mucin 1) and, utilizing machine learning–based modeling, identified a proof-of-concept multiomic model consisting of less than 20 top-performing lipid and protein features. This model was trained on cohort 1 and tested on cohort 2, achieving AUCs of 92% (95% confidence interval, 87%–95%) for distinguishing ovarian cancer from controls and 88% (95% confidence interval, 83%–93%) for distinguishing early-stage ovarian cancer from controls. These findings demonstrate the clinical utility and robustness of lipids as proof-of-concept diagnostic biomarkers for early ovarian cancer within the clinically complex symptomatic population, particularly when applied in a multiomic approach. Significance: Patients with ovarian cancer endure delayed diagnosis and poor outcomes. We profiled lipids in two cohorts and integrated them with proteins in machine learning. This enabled early-stage detection in a complex range of controls.