Using Environmental Mixture Exposure-Triggered Biological Knowledge-Driven Machine Learning to Predict Early Pregnancy Loss.

The assessment of how environmental mixture exposures affect reproductive health faces difficulties. While knowledge graph networks offer valuable advantages in biological interpretation and prediction, their application in epidemiological studies, particularly in a small sample size setting, remains scarce. We recruited 116 women undergoing in vitro fertilization and embryo transfer (IVF-ET) treatment in Beijing and Yantai City, China. Among them, 55 women were diagnosed with early pregnancy loss (EPL), while 61 achieved clinical pregnancy. Clinical records, and paired hair, serum, and follicular samples were collected, with 16 per- and polyfluoroalkyl substances (PFAS) and 41 metal(loid)s measured. We developed a framework coupled with biological knowledge graph-based networks (BKGNs) and machine learning (ML) to predict EPL. Our BKGNs integrate chemical-specific biological pathways, i.e., Gene Ontology (GO) and protein, with individual-level mixture exposure data. The GO-integrated model, with an area under the curve (AUC) of 0.876, outperformed others (AUC = 0.819), even when the sample size decreased to 60% of the total. Additionally, this framework deciphered critical exposures (e.g., serum selenium and chromium) and biological perturbations (e.g., cell population proliferation and apoptotic nuclear changes), linking mixture exposure to EPL. Our proposed novel framework is both robust and cost-effective, offering a mechanistic lens for predicting exposure-associated health outcomes.