Today's Clinical Lab - News, Editorial and Products for the Clinical Laboratory
A pregnant woman hands using a lancet on her finger to check her blood sugar levels with a glucose meter.
The study demonstrates the advantage of using cfDNA sequencing for early detection of GDM and potentially in precision medicine.
iStock, dragana991

Deep Learning Model Noninvasively Predicts Gestational Diabetes Risk

The neural network was built to analyze the genes associated with GDM from cfDNA sequencing data

BGI Genomics
Published:Feb 01, 2024
|2 min read
Register for free to listen to this article
Listen with Speechify

Gestational diabetes mellitus (GDM) is a prevalent pregnancy complication posing significant health risks to both mothers and their newborns. Early detection and treatment of GDM are crucial to prevent adverse outcomes. Current screening methods, like glucose tolerance tests, are in after 24 weeks of pregnancy and have limitations in patient compliance and accuracy.

A new study led by Lijian Zhao, professor, CEO of BGI Genomics, along with Pei Sun, an assistant research fellow, Hui Huang, PhD, and Nan Li, PhD, in collaboration with the Beijing Obstetrics and Gynecology Hospital, recently published on Briefings in Bioinformatics aimed to develop a noninvasive method for early detection of GDM using cell-free DNA (cfDNA) and deep learning models.

Identifying crucial genes in cfDNA using deep learning

A deep learning model was developed using circulating cfDNA samples from 5,085 pregnant women, including 1,942 GDM patients and 3,143 healthy controls, to predict GDM status. The researchers built a Convolutional Neural Networks (CNN)-based deep neural network with a self-attention layer to analyze copy numbers of cfDNA sequencing data linked to GDM, focusing on identifying crucial genetic regions for accurate classification.

The team found that the risk of GDM can be predicted noninvasively in the first trimester, much earlier than the traditional method of glucose tolerance test performed late in pregnancy. The model showed high accuracy (93.5 percent) in predicting GDM status, outperforming traditional methods. However, there’s potential to improve patient compliance and accuracy in GDM screening, providing a promising approach for early detection of GDM.

The study demonstrates the advantage of using cfDNA sequencing for early detection of GDM and highlights the importance of further investigating the use of deep learning models in precision medicine. 

The analysis of cfDNA CNVs associated with diabetes genes and the use of a sophisticated model incorporating a self-attention layer to identify important genetic regions for accurate classification also allowed to identify:

  • Essential genetic regions for accurate classification of GDM.

  • CNV fragments covering 2190 genes, including known genes related to GDM like alpha- and beta-defensin genes (DEFA1, DEFA3, and DEFB1).

  • Enriched biological processes and pathways linked to diabetes, such as glutamate signaling, forebrain development, and GTPase regulator activity.

These findings are a significant leap forward in understanding molecular mechanisms underlying GDM and offering insights for future research and therapeutic strategies.

- This press release was originally published on the BGI Genomics website