Comprehensive Summary
This study, published by Ren et. al., investigates the potential for biological knowledge graph-driven networks (BKGNs) to be used to calculate the risk of early pregnancy loss after exposure to the environmental pollutants Per- and Poly-fluoroalkyls (PFAs) and metalloids, and subsequently predict the underlying biological mechanisms behind the losses. The study was conducted as a prospective cohort study from 2015-17, with 116 participants providing hair, serum, and follicle samples to be digested and assessed for metalloid and PFA content; 55 of these participants experienced early pregnancy loss, while the remaining 61 were able to achieve pregnancy and were thus used as controls. Two different BKGNs - labelled Exposure-Gene Ontology (GO)-Disease (EGD) and Exposure-Protein-Disease (EPD) - were developed from open databases to assess the correlation between pollutant exposure and pregnancy loss, and these were converted to matrices that could delineate an individual's specific phsyiological response to pollutant exposure. There were also attempts to make machine learning models, both with the data from the BKGN and without. The results of the study show that iron (Fe) was present in the highest concentration in serum, zinc (Zn) was present the most in hair, and PFOS and PFOA were the most abundant PFAs overall. Zinc from serum had the strongest correlation with early pregnancy loss, with a Spearman correlation of 6.81, followed by chromium from serum and rubidium in hair at 5.84 and 4.41, respectively. In both the EGD and EPD BKGN, arsenic had the most connections to early pregnancy loss, with 299 and 66 connections respectively. The strongest EGD pathway for early pregnancy loss that was generated was accidental cell death, and the strongest EPD pathways generated were cellular tumor antigen protein 53 (TP53), mitogen activated protein kinase 1 (MAPK1), and interleukin-8 (IL-8). Researchers also determined the accuracy of machine learning models incorporating the data from the BKGNs at predicting early pregnancy loss vs models strictly relying on clinical diagnosis - the respective area under curve (AUC) values were 0.876 and 0.774.
Outcomes and Implications
The use of BKGNs is currently quite rare in epidemiological studies, even though BKGNs - through their integration of results from previous experiments and bioassays - have been shown to be a cost-effective method of determining a variety of biological pathways involved in pregnancy loss, as with the BKGNs, 750 EGD pathways and 159 EPD pathways were generated. Moreover, as evidenced by their greater AUC value, machine learning models that take into account data from BKGNs tend to have improved in their ability to assess the risk of early pregnancy loss. The caveat to note here is that there was a greater difference in accuracy between the machine learning model integrating the Gene Ontology pathways and the non-BKGN machine learning model (as measured by AUC values), than the protein-integrating machine learning model and the same non-BKGN machine learning model. This can most likely attributed to the greater influence of Gene Ontology pathways than proteins in early pregnancy loss, though further studies need to be done regarding this hypothesis. Further studies also need to be done with other pollutants, as the focus of this study on the metalloid and PFA pollutants (the latter of which has not been shown to have a relation with early pregnancy loss) prevents the results of the study from being extrapolated.