PB1520 - Deep learning extracts embedded signatures associated with pregnancy and preeclampsia from circulating vesicle profiles

Tuesday, June 27, 2023

18:30 – 19:30 ET

Room: Exhibition

Presenting Author(s)

Martin Kenny, MSc, PhD

Postdoctoral researcher
University College Dublin
Dublin, Dublin, Ireland

Co-Author(s)

RS

Ravi Shukla

Data Science Consultant
Performance Analytics group (PAG), Dell Technologies
NA, Singapore
Luisa Weiss, BSc, MSc, PhD

Post-doctoral Scientist
School of Biomolecular and Biomedical Science, University College Dublin, Dublin, Ireland
Dublin, Dublin, Ireland
EF

Ella Fouhy

PhD Student
School of Biomolecular and Biomedical Science, University College Dublin
Dublin, Dublin, Ireland
CM

Claire Murphy

Medical Doctor
Department of Haematology, Mater Misericordiae University Hospital and Rotunda Hospital, Dublin, Ireland
Dublin, Dublin, Ireland
FN

Fionnuala Ní Áinle

Consultant Haematologist
Department for Haematology, Mater Misericordiae University Hospital, Dublin, Ireland
Dublin, Dublin, Ireland
Barry Kevane, MB PhD

Consultant Haematologist
Mater Misericordiae University Hospital & University College Dublin
Dublin, Dublin, Ireland
BM

Brian MacNamee

Associate Professor
Insight SFI Research Centre for Data Analytics
Dublin, Dublin, Ireland
JO

John O'Shea

Chief Technologist
Industry Digital Ecosystems, Office of the CTO (OCTO), Dell Technologies
Bratislava, Bratislava, Slovakia
Patricia B. Maguire, PhD

Full Professor
School of Biomolecular and Biomedical Science, University College Dublin, Dublin, Ireland
Dublin, Dublin, Ireland

Background: Extracellular vesicles (EVs) are physiologic mediators of intercellular communication regulating a myriad of biological processes. During pregnancy, EVs exchanged via the blood-placental barrier mediate communication between foetal and maternal systems. Characterization of maternal EV profiles by nanoparticle tracking analysis (NTA) demonstrates observable differences in plasma profiles of pregnant women which distinguish healthy pregnancy from preeclampsia. However, subtle differences in these profiles (figure 1) are difficult to quantify with standard methods, thus limiting the information that can be extracted from these data. This remains an important clinical knowledge gap with mechanistic and therapeutic relevance.

Aims: To develop a deep learning model aimed at differentiating pregnancy phenotypes based on circulating vesicle profiles and to evaluate their potential to detect preeclampsia.

Methods: We possess a unique database of 284 NTA profiles collected from healthy pregnant and non-pregnant women, and pregnant women with either term or early-onset preeclampsia. Utilizing this unique database in combination with patient demographics, we have leveraged the power of deep learning to extract embedded signatures of health and disease in pregnancy.

Results: A convolutional neural network was developed to extract disease-specific signatures (embeddings) from circulating vesicle profiles that can distinguish subtle differences relating to health status (figure 1). The model produced similar embeddings for profiles that were excluded during training and demonstrated improved clustering when patient demographic information was incorporated into the input. Classification models trained on these embeddings were capable of discriminating preeclamptic from healthy pregnancy, regardless of the time of onset.

Conclusion(s): Deep learning can extract signatures of preeclampsia from circulating vesicle profiles to provide insight into disease phenotypes and personalised patient therapeutic responses.