Predictive modelling of prematurity in asymptomatic singleton gestations using maternal plasma proteomics
This repository contains the code and computational work undertaken as part of my (RP2) dissertation submitted to The University of Manchester for the degree of Master of Science in Bioinformatics and Systems Biology in the Faculty of Biology, Medicine and Health.
Premature birth is the leading cause of neonatal mortality and morbidity worldwide. It has multiple aetiologies and clinical phenotypes, making it challenging to comprehend the underlying drivers of phenotype shift from term to preterm during gestation. This project integrated two maternal plasma proteomics datasets to identify differentially expressed proteins, visualise the network topology of term and preterm phenotypes in the second and third trimester via hypergraph analysis, and develop a machine learning classifier to detect term versus preterm phenotype based on maternal plasma protein abundance data. The results identified 459 and 462 differentially expressed proteins in the second and third trimester, respectively, and a significant difference in network topology in the second trimester, compared to the third trimester. The developed machine learning model achieved an AUC of 1.0 (95% CI: 0.996–0.999), showing great potential in future applications. Future directions for the project include development of a hypergraph neural network using the selected features from this model for more informed predictions that account for network topology.
PCA of sample signals demonstrates efficient batch correction across second (left) and third (right) trimesters, with clear
clustering based on the underlying biology. In second and third trimester data, the first principal
component (PC1) explains 17.26% and 27.68% of variance, respectively, whereas the second
principal component (PC2) explains 12.07% and 7.46% of variance. Total
cumulative variance explained by 20 PCs is 61.76%. The PC1 scores were found to differ
significantly across control, pPROM and sPTD groups, meaning that the three
pregnancy outcomes differ across the primary, and strongest, axis of variance in the data –
ultimately depicting that the features (i.e. plasma proteins) that load onto and strongly
contribute to PC1, also known as the first latent variable (LV1), are differentially expressed
across pregnancy outcomes, with the strongest difference occurring between control and sPTD
samples.
A binary classification model has been developed, using a random forest classifier, to assess
the predictive power of plasma protein abundance data in predicting sPTD. Five sets of 50
selected features were used to train separate models for comparison, including i) PC1 loadings,
ii) differentially expressed proteins, iii-iv) proteins with normalised ranks in the 90th and 10th
percentile of the hypergraph, and v) proteins whose normalised ranks are in the 90th percentile
in controls and 10th percentile in sPTD group. In evaluation, this project focused on specificity,
sensitivity and AUC as evaluation metrics due to their high adoption and prevalence in
biomedical and clinical studies, accompanied by MCC for validation.
The cross-validation scores [...] follow a consistent increase until 500 data samples, where they decrease together with the training scores. Despite the different patterns, ROC analysis yielded an AUC of 1.0 (95% CI: 0.996–0.999) and 0.99 (95% CI: 0.989–0.998) for models trained on the PC1 loadings and DE proteins as features, respectively, suggesting excellent discriminative ability against the control and sPTD classes. Both models achieved specificity of 1.0 and sensitivity of 0.99, suggesting a strong capability of accurately identifying positive classes while minimising false positives. The model scoring third by predictive power was trained on the top 50 proteins with the lowest normalised ranks in the 10th percentile of central clusters from the hypergraph analysis. That model achieved an AUC of 0.90 (95% CI: 0.875–0.934), specificity of 0.87 and sensitivity of 0.85, suggesting strong predictive potential among central cluster proteins within the 10th percentile of hypergraph normalised ranks, compared to highly functional proteins in the 90th percentile, which achieved an AUC of 0.5 (95% CI: 0.447–0.557), equivalent to random guessing, demonstrating no evidence of discriminatory ability between sPTD and control groups.
The PC1 and DE-feature models reached an MCC of 0.96, reaching a strong positive correlation between predicted and true group classifications, and reinforcing the previous model evaluation, demonstrating robust and strong predictive abilities of these features. The model trained on the 10th percentile proteins reached an MCC of 0.56, a moderate positive correlation, whereas models trained on the 90th percentile proteins, and proteins that were within the 90th percentile in controls but in 10th percentile in the sPTD group had MCC values of 0, indicating no correlation between predicted and true group classifications.
From the selected protein features from PC1 loadings, DE and 10th percentile proteins used to train the model, 40%, 26%, and 18% were found evolutionarily constrained (LOEUF < 0.6), respectively, whereas no overlap with human accelerated regions was found by genomic ranges encoding the plasma proteins used.
The project repository is structured in the following manner:
002_RP2/
├── hypergraph/
│ ├── hypergraph.Rproj
│ ├── code/
│ ├── data/
│ │ ├── HARs/science.abm1696_table_s1.xlsx
│ │ ├── loeuf/gnomad.v4.1.constraint_metrics.tsv
│ │ ├── hu_et_al_2025/
│ │ │ ├── hu_amnion_obs.csv / _var.csv / _X.csv
│ │ │ └── data_conversion/ (Rmd + amnion.h5ad/.h5Seurat + GSE260715_amnion.RData(.gz))
│ │ └── tarca_et_al_2021/
│ └── output/
│ └── tarca/
├── predictive-modelling/
│ ├── code/
│ ├── data/
│ │ ├── clinical-means-of-preterm-birth-prediction.xlsx
│ │ └── tarca-et-al-2021/
│ └── output/
│ └── tarca_et_al_2021/
└── single-cell-exploration/
├── notebooks/
├── data/
│ ├── arutyunyan-et-al-2023/
│ ├── barrozo-et-al-2023/
│ ├── suryawanshi-et-al-2018/
│ ├── sc-prematurity/
│ ├── vento-tormo-et-al-2018/
│ ├── wang-et-al-2024/
│ └── uhm-et-al-2025/
└── output/
├── ivans/
└── wang/
This project combines a Python workflow (predictive-modelling/) and an R workflow (hypergraph/). The single-cell-exploration/ was an exploratory part conducted at early stages of the project, that did not make it into the results. Package requirements and versions are listed below.
| Package | Version |
|---|---|
| Python | 3.12.1 |
| Jupyter Lab | 4.4.5 |
| pandas | 2.2.3 |
| matplotlib | 3.9.2 |
| numpy | 2.0.0 |
| sklearn | 1.6.1 |
| seaborn | 0.13.2 |
| plotly | 5.24.1 |
| pycombat | 0.3.3 |
| neuroCombat | 0.2.12 |
| harmonypy | 0.0.10 |
| fancyimpute | 0.7.0 |
| scipy | 1.16.0 |
| torch | 2.7.1 |
| torch_geometric | 2.6.1 |
| Package | Version |
|---|---|
| R | 4.4.1 |
| RStudio | 2024.12.1+563 |
| tidyverse | 2.0.0 |
| rstatix | 0.7.2 |
| dendextend | 1.19.1 |
| gplots | 3.2.0 |
| plyr | 1.8.9 |
| ggpubr | 0.6.1 |
| ggsci | 3.2.0 |
| BioQC | 1.34.0 |
| readxl | 1.4.5 |
| GenomicRanges | 1.58.0 |
| GenomicFeatures | 1.58.0 |
| AnnotationDbi | 1.68.0 |
| TxDb.Hsapiens.UCSC.hg38.knownGene | 3.20.0 |
| org.Hs.eg.db | 3.20.0 |
| annotables | 0.2.0 |
| clusterProfiler | 4.14.6 |
| enrichplot | 1.26.6 |
| pathview | 1.46.0 |
| janitor | 2.2.1 |
If you use or refer to this work, please cite it as:
Marčec, D. (2026). Predictive modelling of prematurity in asymptomatic singleton gestations using maternal plasma proteomics [Master's thesis, University of Manchester]. GitHub. https://github.com/doramarcec/RP2
The code in this repository is licensed under the MIT License. The dissertation abstract, text (Main results) and figures are made available under CC-BY-4.0.