PCA and clustering on combined diet and mycobiome profiles

Participant-level PCA of scaled nutrient intake and genus relative abundance, with hierarchical clustering to explore diet–mycobiome groupings.

1 Research question

Can participants be grouped by combined diet (nutrients) and mycobiome (genus taxa) profiles?

Pre-specified method; exploratory interpretation: with 12 participants, PCA and clustering describe patterns in the data but do not provide stable cluster validation (no formal cluster significance testing).

2 Data

Role	Source
Genus taxa	data/processed/genus.csv (make save)
Diet	data/processed/dietary_cleaned.xlsx
Disease group (reference)	Study_group_new on mycobiome metadata

• Unit of analysis: Participant (one row per Participant_ID).
• Mycobiome: Mean genus relative abundance across stool samples per participant.
• Diet: Mean nutrient intake across diary days per participant (same as merge_files.R).
• Features: Panel of nutrients from NUTRIENT_FIELDS plus all genus columns (g__*) with non-zero variance across participants.
• Preprocessing: Features are centered and scaled before PCA and clustering.

Coverage: 12 participants, 51 features (11 nutrients + 40 genera).

3 Methods

Item	Choice
Integration	Left-join participant-mean genus abundances and nutrients on Participant_ID
PCA	stats::prcomp, center = TRUE, scale. = TRUE
Clustering	Hierarchical (Ward.D2) on Euclidean distance of scaled features
Cluster sizes	k = 2 and k = 3 (cutree)
Reference labels	Study_group_new overlaid on PCA (not used to fit clusters)
Notebook reference	notebooks/pca_tsne.Rmd (taxa-only PCA/t-SNE)

# Results objects created in setup: pca_fit, cluster_fit, scores

4 Assumptions and diagnostics

• Sample size: 12 participants; PCA axes and clusters are unstable and sensitive to outliers.
• Scale: Nutrients (g, kcal, mg) and compositional genus proportions are jointly z-scored; relative weighting is arbitrary but standard for integrated PCA.
• Compositional data: Genus abundances are not CLR-transformed here; results complement Bray–Curtis / PERMANOVA analyses on taxa alone.
• Independence: Technical stool replicates are averaged before analysis.

5 Results

5.1 Variance explained

PC	Component	Variance (%)	Cumulative (%)
PC1	1	16.08	16.08
PC2	2	14.71	30.79
PC3	3	12.25	43.04
PC4	4	11.66	54.71
PC5	5	9.85	64.56
PC6	6	8.72	73.28
PC7	7	7.88	81.16
PC8	8	6.61	87.76
PC9	9	5.18	92.94
PC10	10	3.85	96.80

PC1 and PC2 explain 16.1% and 14.7% of variance (cumulative PC1–3: 43.0%).

5.2 Clusters vs study group

Participant counts: hierarchical clusters (k = 2) vs study group

cluster	Active IBD	Non-IBD	Quiescent
1	3	4	2
2	1	0	2

Participant counts: hierarchical clusters (k = 3) vs study group

cluster	Active IBD	Non-IBD	Quiescent
1	2	3	2
2	1	0	2
3	1	1	0

Clusters are unsupervised and need not align with clinical study group. Compare tables and PCA colours to see whether diet–mycobiome structure tracks disease status or forms separate groupings.

6 Figures

6.1 Cumulative variance (scree)

plot_pca_variance(pca_fit$variance_df, max_pc = min(10L, n_participants))

Figure 1: Cumulative proportion of variance explained by principal components.

6.2 PCA scores (PC1 vs PC2)

plot_pca_scores(
  scores,
  colour_col = "Study_group_new",
  variance_df = pca_fit$variance_df,
  title = NULL
)

Figure 2: PCA of combined nutrient and genus features. Points are participants; colour is study group.

6.3 PCA with hierarchical clusters (k = 2)

plot_pca_scores(
  scores,
  colour_col = "Study_group_new",
  shape_col = "k2",
  variance_df = pca_fit$variance_df,
  title = "PCA with hierarchical clusters (k = 2)"
)

Figure 3: Same PCA space with point shape indicating Ward clusters (k = 2).

6.4 Top PC1 loadings

plot_top_pca_loadings(pca_fit$loadings, pc = "PC1", n_top = 10L)

Figure 4: Features with largest absolute loadings on PC1 (nutrients and genera).

6.5 Clustering dendrogram

plot_cluster_dendrogram(cluster_fit)

Figure 5: Ward hierarchical clustering dendrogram on scaled diet + genus features (participant IDs as leaves).

7 Interpretation

PCA reduces the joint nutrient + genus feature space to orthogonal axes for visualization. If participants with similar diets and mycobiomes sit near each other, they may form visual groups; formal cluster validity was not assessed (silhouette, gap statistic, etc.) given small n.

Hierarchical clusters partition participants without using Study_group_new. Agreement between clusters and clinical group supports an integrated diet–mycobiome signal; disagreement suggests heterogeneity within disease labels or driven by diet/taxa not captured by group alone.

8 Reproducibility

• Helpers: stats/R/pca_clustering_helpers.R, stats/R/nutrient_association_helpers.R, stats/R/permanova_helpers.R
• Notebook: notebooks/pca_tsne.Rmd (taxa-only PCA/t-SNE)
• Render: quarto render stats from repository root