Skip to main content
Premium Trial:

Request an Annual Quote

Preterm Birth Biology Unraveled With Multiomics, Machine Learning

NEW YORK – An international team led by investigators at Stanford University School of Medicine has unearthed maternal features associated with preterm birth in several low- and middle-income countries using a combination of multiomics and machine learning.

The work builds on previous research highlighting a range of epidemiological factors — ranging from body mass index (BMI), socioeconomic status, or environmental exposures to short intervals between pregnancies, previous delivery by cesarean section, or a history of adverse pregnancy outcomes — that can influence an individual's preterm birth risk.

"Preterm birth is the single largest cause of death in children under 5, both in low- and high-income countries," co-senior and corresponding author Nima Aghaeepour, a researcher at Stanford and vice chair of research in data science, said in an email.

"[U]sing machine learning and multiomics analysis, we have demonstrated that … epidemiologic factors don't belong in the same 'bucket,'" Aghaeepour explained. "They each impact biology differently, and likely have their own specific influence on the biology for prematurity."

As they reported in Science Advances on Wednesday, the researchers brought together maternal data for more than 13,800 samples from pregnant participants at five sites in Bangladesh, Zambia, Pakistan, and Tanzania. They turned to liquid chromatography-mass spectrometry for untargeted metabolomic and targeted lipidomic analyses on mid-pregnancy blood plasma samples for a subset of 231 individuals. Multiplex proteomic profiling on nearly 1,200 proteins in the blood samples was performed at Olink Proteomics.

Together with clinical profiles for the maternal participants, and machine learning-based preterm birth prediction models based on various clinical and epidemiological features, the multiomic profiles provided a biological perspective on the epidemiological risk factors linked to preterm birth in this and previous studies.

The results suggested that fetal and placental proteins such as placental alkaline phosphatase (ALPP), alpha-fetoprotein (AFP), and placental growth factor (PGF) tend to coincide with time-to-delivery and the presence of certain pregnancy-related steroid hormones, for example.

Time-to-delivery also tracked with certain immune proteins such as the programmed death ligand-1 (PD-L1), C-C motif chemokine ligand 28 (CCL28), and the leukemia inhibitory factor receptor (LIFR), but was negatively correlated with levels of a metabolite called cortolone glucuronide.

The team found that increasing maternal age was linked to lower-than-usual levels of the type IX collagen protein COL9A1, on the other hand. Meanwhile, levels of endothelial nitric oxide synthase (eNOS) and the inflammatory chemokine CXCL13 tracked with gravidity, or the number of pregnancies an individual has had.

Leptin, structural protein fatty acid binding protein 4 (FABP4), lipid-associated perilipin 1 (PLIN1), and certain metabolic pathway contributors appeared to be positively associated with maternal BMI, the researchers reported, while other compounds had negative ties to BMI. In the current analysis, preterm birth tended to be more common in individuals with lower-than-usual BMI.

"Our study provides an integrated view of how previously identified epidemiological associations such as parity [number of previous births], BMI, and age are etiologically connected to an underlying biology evidenced by various 'omic correlates," the authors reported. "This integrated view was possible by leveraging a large, multinational cohort together with multiomic profiling in a subcohort to simultaneously explore the population-level correlates of [preterm birth] and how these associate with individual women's biology."

Consequently, the authors suggested that the study may help in improving the performance of preterm birth prediction models, which are typically based on population-level data that can be difficult to apply to individuals, thus improving the likelihood that related interventions may be more broadly beneficial.

"This is a first step towards precision medicine for these patients," Aghaeepour said.

"Neither biology nor epidemiology is going to address this disease in isolation," he suggested. "We need to bring these two fields together, and machine learning is the perfect channel for enabling these two otherwise separate fields to communicate with each other."