Multi-Omics Pathway-based Data Integration Using Machine Learning and Large Language Models (LLMs)
Five-month individual MRes research project, Department of Systems Medicine, Imperial College London, supervised by Prof. Timothy Ebbels with Dr Cecilia Wieder, in collaboration with Erik Huckvale and Prof. Hunter Moseley at the University of Kentucky.
The question
PathIntegrate (Wieder et al., PLOS Computational Biology 2024) converts each omics layer into single-sample pathway scores, so proteomics and metabolomics can be integrated and modelled on one interpretable scale. Metabolomics is its weak point, for two reasons that compound each other: many metabolite names never map to a database identifier at all, and curated pathway databases list far more proteins than metabolites. A pathway method is only as good as the fraction of measured molecules it can place.
The data was the Su et al. 2022 COVID-19 multi-omics cohort: Olink proteomics and Metabolon untargeted metabolomics, 128 patients with both layers (45 mild, 83 severe) after quality control, imputation and filtering, leaving 933 metabolites and 455 proteins.
Aim A: mapping more metabolites with semantic search
Roughly 400,000 ChEBI compound names were embedded with the all-mpnet-base-v2 sentence transformer and indexed in Elasticsearch, so a query name is matched by nearest neighbour in embedding space rather than by exact string. I built a hand-curated gold standard of ChEBI IDs for 713 of the 933 metabolites, checked by Dr Wieder, and scored precision, recall and AUC across similarity cut-offs. A Streamlit app, “Find Your Metabolite ChEBI ID!”, wraps the search.
At the 0.75 cut-off where precision and recall cross (both about 0.87), the search returned 591 unique IDs reaching 111 pathways, against MetaboAnalyst’s 417 IDs reaching 70 pathways.
Classification of mild against severe was comparable rather than better: AUC 0.86 for MetaboAnalyst, 0.81 for the first LLM version and 0.87 once ties were resolved in favour of KEGG-sourced IDs. The first version’s dip traced to duplicate ChEBI entries from different sources, with 110 metabolites having more than one exact match. The honest summary is similar accuracy over much wider coverage.
Aim B: augmenting KEGG with predicted pathway links
I reproduced Huckvale et al.’s multilayer perceptron for metabolite-pathway prediction, about two weeks of training on Imperial’s HPC. It covered only 12 broad KEGG categories, so I proposed extending it to the 172 individual metabolic pathways; the Kentucky group ran that extension and returned predictions with a Matthews correlation coefficient per pathway.
Filtering those predictions at MCC > 0.7 and rebuilding the database took KEGG from 288 pathways covering 3,192 metabolites to 344 pathways covering 10,959 metabolites, a 3.4-fold increase. In the COVID-19 data the metabolites reached 154 pathways instead of 104. Classification AUC was 0.931 against 0.919, a small gain given the spread.
A caution worth stating plainly: these are model predictions filtered by a quality threshold, not curated annotations. Converting between identifier systems is also less reliable than it looks: three standard tools agreed on only 35 of 315 KEGG IDs.
Aim C: unsupervised PathIntegrate
PathIntegrate was supervised only, so I forked it and added two modules operating on the integrated pathway matrix: SingleViewDimRed (PCA with biplots and pathway loadings, t-SNE) and SingleViewClust (KMeans, agglomerative and BIRCH clustering, consensus clustering over subsampled runs, silhouette-based selection of cluster number, ARI against known labels).
In pathway space, PCA separated mild patients from the most severe more clearly than the molecular data did, with Innate Immune System among the pathways loading on PC1. Clustering recovered the mild group but mixed the two severe groups (ARI 0.25). On simulated data with planted pathway enrichments, all three algorithms recovered two clusters perfectly and degraded to ARI 0.2 to 0.27 by seven or eight, with no algorithm winning at every k.
Outcome
The unsupervised module was released in PathIntegrate v1.0.0 (Zenodo, creators Popham, Wieder and Ebbels) and ships in the package on PyPI, whose feature list now leads with “NEW unsupervised SingleView models (dimensionality reduction and clustering in the pathway space)”. Source is at cwieder/PathIntegrate, with a tutorial in MultiomicsML. PathIntegrate and ssPA are Dr Wieder’s work; the contribution here is the unsupervised extension.