Multi-omic profiling and machine learning reveal diagnostic metabolic signatures and distinct decidual cellular states in tubal pregnancy and spontaneous abortion
This study demonstrates that metabolic profiling and machine learning can potentially improve the early diagnosis of tubal pregnancy and spontaneous abortion, but further validation in larger cohorts is necessary.
Where it sits
this study against the rest of the glutathione (gsh) corpusSummary and findings
This study measured metabolic signatures in serum samples from patients with tubal pregnancy (TP), spontaneous abortion (SA), and normal pregnancy (NP). Two machine learning models were developed to identify TP and SA based on metabolic perturbations. Key biomarkers identified included glutathione for SA prediction and 15-deoxy-Δ¹²,¹⁴-prostaglandin J₂ for TP discrimination.
Abstract
<title>Abstract</title> <p> <bold>Background</bold> Tubal pregnancy (TP) and spontaneous abortion (SA) are prevalent early pregnancy complications that severely compromise gestational safety. Accurate differentiation between normal intrauterine pregnancy (NP), TP and SA at early gestational stages is critical for standardized clinical management and the reduction of adverse pregnancy outcomes. Clinically, intrauterine pregnancy (IUP) includes both normal viable intrauterine pregnancy and non-viable intrauterine pregnancy such as early spontaneous abortion, which serves as the main control cohort for distinguishing ectopic gestation. Traditional diagnostic approaches relying on serial β-hCG detection and transvaginal ultrasound exhibit limited accuracy for both pregnancy localization and embryonic viability assessment, easily resulting in delayed or inappropriate clinical intervention for TP and SA patients. <bold>Methods</bold> Untargeted serum metabolomics based on UHPLC-MS/MS was performed on serum samples from 50 patients with TP, 50 patients with SA and 50 individuals with uncomplicated NP. Multivariate statistical analyses and KEGG pathway enrichment were used to screen differential metabolites and characterize disease-specific metabolic perturbations. Samples were randomly allocated to a training set comprising 60% of the cohort and a validation set comprising the remaining 40%. Diagnostic models for pregnancy localization discrimination and embryonic viability assessment were established, with model efficacy verified via multiple indicators and core predictive metabolites screened through SHAP analysis. Furthermore, scRNA-seq of decidual tissues from TP, SA and NP patients was performed to decipher the cellular and molecular mechanisms underlying disease-specific metabolic perturbations. <bold>Results</bold> Two machine learning models were constructed for the staged identification of abnormal early pregnancy. First, metabolic perturbations of TP were characterized by comparing TP with combined IUP (NP + SA) samples. In total, 77 differential metabolites and nine disturbed metabolic pathways were identified, which were mainly involved in arachidonic acid-related inflammation, TCA cycle dysfunction and tryptophan metabolic disorder. The LightGBM model based on LASSO-selected core metabolites showed valid performance in TP identification (AUC = 0.948, PR-AUC = 0.919, accuracy = 0.90). 15-deoxy-Δ¹²,¹⁴-prostaglandin J₂ and kynurenine were identified as key biomarkers for TP discrimination. Second, metabolic signatures distinguishing SA from NP were screened to evaluate embryonic developmental status. A total of 81 differential metabolites were obtained in SA samples, which were enriched in lipid peroxidation and ferroptosis-associated pathways. The LightGBM model for SA identification achieved valid predictive performance (AUC = 0.985, PR-AUC = 0.986, accuracy = 0.95). Glutathione and 3-hydroxyanthranilic acid were core biomarkers for SA prediction. Single-cell transcriptomic analysis further identified decidual cell states concordant with systemic serum metabolic alterations. TP decidua was characterized by a TP-enriched macrophage state with reprogrammed lipid inflammatory and energy metabolism profiles. SA decidua exhibited pro-inflammatory macrophage polarization, accompanied by downregulated <italic>GPX4</italic> expression and elevated ferroptosis-associated transcriptional signatures in endothelial cells and pro-inflammatory macrophages. These cell-type-specific transcriptional alterations were concordant with, and may provide a cellular context for, the systemic metabolic perturbations observed in serum. <bold>Conclusions</bold> This study combines serum untargeted metabolomics and uterine decidual single-cell transcriptomics to characterize divergent metabolic dysregulations in TP and SA. The LightGBM metabolite-based models enable reliable early discrimination of ectopic pregnancy and impaired embryonic viability, offering potential auxiliary biomarkers for risk stratification in early pregnancy. </p>
Background
This paper addresses the challenge of differentiating between normal intrauterine pregnancy, tubal pregnancy, and spontaneous abortion at early gestational stages. Traditional diagnostic methods, such as serial β-hCG detection and transvaginal ultrasound, have limitations in accuracy. The study aims to improve diagnostic accuracy through metabolic profiling and machine learning techniques.
Methods
The study utilized untargeted serum metabolomics based on UHPLC-MS/MS on samples from 150 patients (50 with TP, 50 with SA, and 50 with NP). Metabolites were analyzed using multivariate statistical methods and KEGG pathway enrichment. Two machine learning models were constructed, with 60% of samples used for training and 40% for validation.
Results
The primary endpoint for TP identification achieved an AUC of 0.948, PR-AUC of 0.919, and accuracy of 0.90. A total of 77 differential metabolites were identified in TP, while 81 differential metabolites were identified in SA. For SA identification, the LightGBM model achieved an AUC of 0.985, PR-AUC of 0.986, and accuracy of 0.95.
Interpretation
The findings suggest that metabolic profiling can significantly enhance the differentiation of early pregnancy complications compared to traditional methods. While the statistical significance of the results is notable, the clinical relevance remains to be established in larger, diverse populations. Confounding factors include the small sample size and the reliance on machine learning models that may not generalize well.
Key findings
- AUC=0.948, PR-AUC=0.919, accuracy=0.90 for TP identification using LightGBM model.
- A total of 77 differential metabolites and nine disturbed metabolic pathways identified in TP.
- AUC=0.985, PR-AUC=0.986, accuracy=0.95 for SA identification using LightGBM model.
- A total of 81 differential metabolites identified in SA samples.
- Glutathione and 3-hydroxyanthranilic acid were core biomarkers for SA prediction.
Limitations
- small sample size of 150 patients
- machine learning models require external validation
- results based on metabolomic and transcriptomic analyses
- not directly linked to clinical outcomes