Text phrase‐mining in identifying and classifying maternal proteins and genes across preeclampsia and similar pathologies

Jacqueline G. Urdang, Stephanie Masters, Nneoma Edokobi, Chitra Mukherjee, Arnib Quazi, David A. Liem, Monica Ahrens, Xuan Wang, Megan WhithamPhysiological 2025 · DOI

TL;DR

This study leverages text phrase-mining and natural language processing (NLP) to efficiently identify and classify maternal proteins and genes (MPGs) associated with preeclampsia and related cardiovascular disorders. By analyzing a vast corpus of obstetrics literature, the researchers developed a scoring system to reveal novel and confirm existing MPG-disease relationships, paving the way for faster biomarker discovery.

What Problem Does This Paper Address?

The field of obstetrics is grappling with an overwhelming volume of published research, making it challenging for clinicians and researchers to stay abreast of the latest findings and identify critical relationships between biological factors and diseases. Specifically, understanding the complex interplay of maternal proteins and genes (MPGs) in conditions like preeclampsia and its overlapping cardiovascular pathologies is a significant hurdle. Manual literature review is time-consuming and prone to human error. This paper addresses the need for automated, efficient methods to mine this literature, uncover hidden connections, and identify potential biomarkers for these critical maternal health issues.

Methods

The core of this study is a text phrase-mining and NLP pipeline, building upon a previously developed system called Context-Aware Semantic Online Analytical Processing (CaseOLAP). The process involved:

  1. Disease Selection: Six diseases with overlapping clinical presentations and relevance to maternal health were chosen: preeclampsia (PE), HELLP syndrome, acute fatty liver of pregnancy (AFL), thrombotic angiopathies (TA), antiphospholipid syndrome (APL), and systemic lupus erythematosus (SLE).
  2. Corpus Generation: Medical Subject Headings (MeSH) terms were used to query PubMed, retrieving 184,235 abstracts related to the selected diseases.
  3. MPG Identification: Expert physicians identified a list of 254 maternal proteins and genes (MPGs) of interest. From these, 44 were found to be mentioned within the retrieved abstracts.
  4. CaseOLAP Scoring: The CaseOLAP algorithm was applied to score the association between each of the 44 MPGs and the six diseases. This score is a product of three components:
  5. Integrity: Measures the coherence and meaningfulness of semantic units within the text.
  6. Distinctiveness: Assesses the specificity of an MPG to a particular disease, with higher scores indicating less overlap across diseases.
  7. Popularity: Reflects how frequently an MPG appears in the literature related to a specific disease.
  8. Statistical Analysis: The scores were analyzed using R, including descriptive statistics, heat maps, count plots, bar plots, and Principal Component Analysis (PCA) to visualize MPG-disease relationships and similarities.

Datasets

  • PubMed Abstracts: A corpus of 184,235 abstracts related to the six selected obstetrics and cardiovascular diseases was used. Specific counts for each disease were: PE (34,925), HELLP (1,960), AFL (43,410), TA (24,709), APL (9,092), and SLE (69,280).
  • Maternal Proteins and Genes (MPGs): An initial list of 254 MPGs was curated by expert physicians, from which 44 were identified as relevant to the study's corpus.

Key Insights and Ideas

  • Automated Discovery of MPG-Disease Links: The study demonstrates that text mining can efficiently uncover and validate complex relationships between maternal proteins/genes and specific disease states, significantly accelerating the discovery process compared to manual methods.
  • Identification of Disease-Specific MPG Signatures: The analysis revealed that while some MPGs are shared across multiple conditions (e.g., those involved in angiogenesis, hemodynamic regulation, metabolism, and inflammation), others show distinct associations with individual diseases, suggesting unique underlying molecular pathways.
  • PIGF as a Potential Preeclampsia Biomarker: Placental growth factor (PIGF) emerged as a protein with a strong relationship to preeclampsia (PE), aligning with existing research and highlighting its potential as a diagnostic or prognostic marker.
  • vWF and Complement in Thrombotic Angiopathies: Von Willebrand factor (vWF) and complement (particularly C5b-9) were strongly associated with thrombotic angiopathies (TA), reinforcing their known roles in hemostasis and immune response in these conditions.
  • MPO and IL-10 in SLE: Myeloperoxidase (MPO) and Interleukin-10 (IL-10) showed strong associations with systemic lupus erythematosus (SLE), suggesting their potential involvement in the inflammatory processes of SLE, particularly in the context of pregnancy.

Key Results

  • 44 MPGs Identified: Out of 254 MPGs of interest, 44 were identified as having associations with the six studied diseases within the PubMed abstracts.
  • Heterogeneous MPG-Disease Relationships: The CaseOLAP scores and subsequent PCA revealed a heterogeneous distribution of MPG-disease relationships, indicating varied biological functions and pathways across the different conditions.
  • Overlapping vs. Distinct Associations: The study confirmed known overlaps in MPG associations for some diseases (e.g., HELLP and TA sharing many MPGs) while also highlighting distinct protein associations for others (e.g., TA and PE).
  • PCA Shows Divergence: Principal Component Analysis (PCA) demonstrated that the six disease states had separated vectors, suggesting disparate MPG scoring patterns and potentially distinct biological underpinnings.

Limitations and Future Work

  • Exclusion of Non-Textual Data: The study exclusively used text abstracts, which may lead to the loss of valuable information present in other data formats (e.g., imaging, numerical data). This approach, however, was chosen to reduce semantic complexity and minimize inaccurate relationships.
  • Inclusion of Review Abstracts: The inclusion of review article abstracts might artificially inflate the "popularity" score of certain MPGs.
  • Lack of Directionality: The CaseOLAP score does not indicate whether a protein is up- or down-regulated in a disease, which could impact the interpretation of relationships.
  • Popularity Score Bias: A low popularity score might not necessarily indicate a lack of relationship but could reflect limited prior research or fewer publications on that specific MPG-disease link.

Future Work: The authors suggest extending this approach to patient-level physician notes for a more granular analysis of health features and disease presentations. Further research could also explore other disease states of interest in obstetrics, such as premature rupture of membranes (PROM), preterm labor (PTL), and maternal mortality.

Summary updated

Abstract

This study aims to demonstrate that text phrase-mining and natural language processing (NLP) can annotate huge quantities of obstetrics textual data for the discovery and evaluation of maternal protein/gene (MPG)-disease interactions involved in the preeclampsia pathway. We employ a phrase-mining/NLP pipeline to evaluate unique MPGs involved in six cardiovascular derangements with overlapping presentations during pregnancy. The diseases were matched with Medical Subject Headings. A textual corpus was developed from abstracts matched to these terms through PubMed. Fourty-four MPGs were identified with respect to the diseases. Processing was performed, with unique scores for each MPG-disease pair. Components of the score were calculated and weighted for distinctness, integrity, and popularity. Statistical analyses were conducted for the examination of protein-disease relationships. Fourty-four MPGs with known associations to cardiovascular disease and preeclampsia pathways were identified among the 6 diseases. MPGs shared across the greatest number of disease states were implicated in: (1) angiogenesis and vasoconstriction, (2) hemodynamic regulation, (3) hormonal regulation of metabolism, and (4) inflammation. NLP and text phrase-mining are successfully applied to Obstetrics abstracts with accuracy and speed. This approach holds promise in synthesizing large volumes of data for presenting trends in the Obstetric literature and for the identification of promising biomarkers.
BibTeX
@inproceedings{urdang2025text,
  author = {Jacqueline G. Urdang and Stephanie Masters and Nneoma Edokobi and Chitra Mukherjee and Arnib Quazi and David A. Liem and Monica Ahrens and Xuan Wang and Megan Whitham},
  title = {{Text phrase‐mining in identifying and classifying maternal proteins and genes across preeclampsia and similar pathologies}},
  booktitle = {Physiological},
  year = {2025},
  doi = {10.14814/phy2.70262},
}
Figure