P-341 Identification of potential biomarkers for peritoneal endometriosis from the whole blood transcriptome by machine learning

Abstract Study question Can blood biomarker candidates for peritoneal endometriosis be identified by an untargeted transcriptomic approach using whole genome RNA sequencing of whole blood and machine learning? Summary answer A machine learning feature selection technique identified six transcripts as potential blood biomarkers for peritoneal endometriosis that need validation in a larger cohort. What is known already Endometriosis, which affects up to 10% of women, occurs when endometrial-like tissue grows outside the uterus. Peritoneal endometriosis (PE), the most common form, is challenging to diagnose as it cannot be detected with standard imaging, with laparoscopy and histological confirmation being the gold standard. This leads to significant delays in diagnosis. Biomarkers could speed up diagnosis and increase accessibility. Despite efforts to identify non-invasive biomarkers, none have met clinical criteria for sensitivity and specificity. Whole genome RNA sequencing offers a more comprehensive, unbiased approach for discovering potential biomarkers, unlike traditional hypothesis-driven methods. Study design, size, duration This study recruited 48 patients from the University Medical Centre Ljubljana (Slovenia, Jan 2020-Dec 2022) and the Medical University of Vienna (Austria, Nov 2019-Jan 2020). Blood samples were collected, and whole genome RNA sequencing was performed. Differentially expressed genes (DEGs) and transcripts (DETs) were identified (FDR<0.05). RNA sequencing data were processed through a machine learning pipeline to identify key genes and transcripts, which were used to develop support vector machine classifiers for predicting endometriosis status. Participants/materials, setting, methods Patients with endometriosis symptoms undergoing laparoscopic surgery were enrolled and categorized into three groups: PE (n = 20), PE + ovarian endometriosis (OE) (n = 8), and controls (n = 20). Principal component analysis (PCA) and feature selection were performed using transcripts per million data. Key genes and transcripts were identified through mutual information, random forest, and SVM feature selection. These features trained an SVM classifier, which was tested on a separate dataset to evaluate its performance. Main results and the role of chance No DEGs were identified in the proliferative phase. In the secretory phase, 48 DEGs were found between all cases and controls, 1,035 between controls and the PE group, and 3 between controls and the PE+OE group. Regarding DETs, 2 were identified between controls and the PE group in the proliferative phase. In the secretory phase, 110 DETs were found between all cases and controls, 922 between controls and the PE group, and 29 between controls and the PE+OE group. There was no common DEGs and DETs in both menstrual phases. PCA clustered samples only based on their menstrual phase. No PCA clustering was observed based on endometriosis status or other metadata. The genes and transcripts identified through feature selection were tested on the test dataset using SVM models. A set of 3 genes achieved the highest ROC AUC (AUC=0.88, sensitivity=0.75, specificity=1.0) in distinguishing controls from cases in the secretory phase. For transcripts, a set of six selected transcripts delivered the highest performance in SVM models on the test dataset, with an ROC AUC of 0.92, sensitivity of 0.75, and specificity of 1.0. Limitations, reasons for caution The main limitations of the study are the relatively small sample size of patients, the exclusion and inclusion criteria for participants, who therefore do not represent the general population. The validation of identified set of DETs is currently underway. Wider implications of the findings The study has uncovered DEGs not previously associated with endometriosis that could help to elucidate the still unexplained pathophysiology. A set of new potential biomarker transcripts for peritoneal endometriosis were identified by calculating mutual information and SVM-RFE. Trial registration number No