Multimodal Integration for Colorectal Cancer

Riya Shet

A reproducible multi-omics pipeline that fuses blood metabolomics, clinical biochemistry, and diet to classify colorectal cancer. It compares two integration strategies and separates what fusion adds for prediction from what it adds for biology.

Python · R (mixOmics) · SHAP · UK Biobank, n = 4,596

Circos plot of cross-block correlations between metabolomics and biochemistry features

The question

A cancer cohort carries several kinds of data per patient. Two questions follow from that. Does combining them classify colorectal cancer better than the best single source? And does combining them reveal biology that no single source shows on its own? I built one pipeline to answer both, and the answers pull in different directions.

Results

Clinical biochemistry alone already does most of the work. Fusing the modalities adds a real but small amount on top, which is the honest headline of this project.

ModelStrategyAUC
Random Forest, biochemistry aloneper-modality0.824
Random Forest, metabolomics aloneper-modality0.565
Stacked logistic regressionlate fusion0.827
rCCA, component 1 (shared structure)intermediateρ = 0.884
ROC curve for the stacked late-fusion classifier with area under the curve near 0.83

Figure 1. ROC for the stacked late-fusion classifier (AUC ≈ 0.83), evaluated with 5-fold cross-validation repeated three times. Late fusion adds roughly +0.004 AUC over biochemistry alone.

What the model uses

I used SHAP and a small surrogate tree to see what drives the decision. Both point to the same place: the biochemistry block, led by HDL, Apolipoprotein A-I, and albumin. The metabolomics block refines the call rather than leading it, which is worth knowing before anyone trusts the output.

SHAP beeswarm plot for the biochemistry random forest, ranking features by impact on the prediction

Figure 2. SHAP values for the biochemistry model. Lipoprotein-A, direct bilirubin, and HDL-cholesterol carry the most weight in the per-patient decisions.

Depth-4 surrogate decision tree approximating the stacked classifier, splitting almost entirely on the biochemistry probability

Figure 3. A depth-4 surrogate tree approximating the stacked classifier. Almost every split is on the biochemistry probability, which makes the biochemistry-led behaviour explicit.

Shared biology

The unsupervised side of the pipeline (DIABLO and regularised CCA) and a cross-modality correlation network all recover one axis. HDL and Apolipoprotein A-I in the clinical block link to HDL particle subfractions in the metabolite block. That shared structure, rather than the prediction score, is what integration actually buys here.

Cross-modality correlation network linking biochemistry HDL features to metabolomics HDL particle subfractions

Figure 4. Cross-modality network at significant Spearman correlations. The dominant bridge runs along the HDL and Apolipoprotein A-I axis shared by both data blocks.

An honest limitation

Fusion adds about +0.004 AUC over biochemistry alone, which is small. The value of integration here is biomarker discovery, the shared HDL and ApoA-I axis, not a better prediction. The data are cross-sectional, so this is association rather than a screening test. The full pipeline, the hyperparameter grid, and the verified clinical markers (every PMID checked on PubMed) are in the repository.