🧬 Final Year Project — University of Peradeniya

Data-Driven Methods For
Comparative Metagenomics

Systematic benchmarking of dimensionality reduction techniques and GNN-based clustering specifically optimised for high-dimensional, sparse, compositional metagenomic data.

⭐ View Repository Learn More ↓

Overview

Project Abstract

Bridging the gap in metagenomics analysis through data-driven methods

Metagenomic studies generate high-dimensional, sparse, and compositional datasets that challenge traditional analytical methods. This project systematically benchmarks 8 dimensionality reduction (DR) methods across 3 diverse metagenomic datasets (Human Gut, Ocean, Potato Soil) and validates Graph Neural Network (GNN) architectures for unsupervised microbial community clustering. Our goal is to provide evidence-based recommendations for the metagenomics community and develop a unified preprocessing-to-evaluation pipeline.

⚡ Extreme Sparsity (90–99% zeros) 🔗 Compositionality Constraints 🌿 Phylogenetic Structure

DR Methods Compared

Diverse Datasets

24+

Visualizations Generated

Research

Research Phases

A two-phase approach to comparative metagenomics analysis

Phase 1

📊 DR Benchmarking

✓ Completed

Datasets: Human Gut (3.6K samples), Ocean (139 samples), Potato Soil (885 features)
Methods: PCA, UniFrac PCoA, MDS, t-SNE, UMAP, PaCMAP, PHATE, SONG
Metrics: Trustworthiness, Continuity, UniFrac/Bray-Curtis/Aitchison correlations
Finding: No universal best method — performance is dataset & metric dependent

Phase 2

🔬 GNN Clustering

◉ In Progress

Graphs: KNN cosine graph built on nzCLR-transformed features
Architectures: DMoN, MinCutPool + K-Means across 8 DR embeddings
Evaluation: NMI, ARI, Silhouette score, Stability analysis
Goal: Validate GNN clustering vs. traditional methods on 3 real metagenomic datasets

Methodology

Research Pipeline

End-to-end workflow from raw data to actionable insights

🧹

Preprocessing

nzCLR transformation & Jaccard/Bray-Curtis distance computation to handle compositionality

📐

Dimensionality Reduction

8 methods: PCA, PCoA, MDS, t-SNE, UMAP, PaCMAP, PHATE, SONG

🕸️

GNN Clustering

KNN cosine graph construction & GNN architectures (DMoN, MinCutPool) + K-Means

📈

Evaluation

Trustworthiness, Continuity, NMI, ARI, Silhouette & cross-dataset correlation analysis

Findings

Key Results

Highlights from Phase 1 DR benchmarking across 3 datasets

0.927

Trustworthiness

Best: SONG method

0.999

UniFrac Correlation

Best: PCoA method

24+

Visualizations

3 datasets × 8 methods

💡 Key Insights

No single DR method universally outperforms all others across every dataset and metric
UniFrac PCoA excels at preserving phylogenetic distance structure (correlation up to 0.999)
PaCMAP and SONG achieve the best balance between local and global structure preservation
GNN-based clustering (KNN+DMoN) captures microbial groupings beyond what flat K-Means achieves

Achievements