Evaluation design

Metrics

The benchmark evaluates latent representations with 20 metrics across clustering quality, projection preservation, and latent-space structure. These are the same metric families used in the manuscript.

Metric definitions are public in preview mode; aggregate result figures remain gated until journal acceptance.

Datasets

27 cancer and 26 development cohorts

Study tracks

7 comparative, 4 robustness and efficiency

Metrics

clustering, DRE, and LSE families

Primary latent

10D

shared dimensionality for learned methods

Axis 01

Clustering

Partition agreement and separation against the shared Leiden reference.

Axis 02

Projection fidelity

Co-ranking measures for UMAP and t-SNE views of each learned latent space.

Axis 03

Latent structure

Intrinsic spectral and geometric diagnostics before plotting or clustering.

Definitions

Metric inventory

Each metric lists its intended direction so reviewers can distinguish optimization targets from diagnostic quantities.

Clustering quality

Agreement and separation of K-means clusters in latent space against the Leiden reference partition.

6 metrics

Metric	Direction	Definition
NMI	higher	Normalized mutual information for partition agreement.
ARI	higher	Adjusted Rand index with chance correction.
ASW	higher	Average silhouette width for intra- versus inter-cluster distance.
DAV	lower	Davies-Bouldin index; lower values indicate less cluster overlap.
CAL	higher	Calinski-Harabasz score for compact, well-separated clusters.
COR	diagnostic	Mean absolute inter-dimensional Pearson correlation in the latent space.

Dimensionality reduction evaluation

Co-ranking evaluation of how UMAP and t-SNE projections preserve neighborhoods from the learned latent space.

8 metrics

Metric	Direction	Definition
UMAP distance correlation	higher	Rank-distance agreement for UMAP projections.
UMAP Q_local	higher	Local nearest-neighbor preservation at k = 15.
UMAP Q_global	higher	Global structure preservation in the projection.
UMAP overall	higher	Combined local and global UMAP quality.
t-SNE distance correlation	higher	Rank-distance agreement for t-SNE projections.
t-SNE Q_local	higher	Local nearest-neighbor preservation at k = 15.
t-SNE Q_global	higher	Global structure preservation in the projection.
t-SNE overall	higher	Combined local and global t-SNE quality.

Latent space evaluation

Intrinsic spectral and geometric diagnostics for the latent representation before 2-D plotting.

6 metrics

Metric	Direction	Definition
Manifold dimensionality	diagnostic	Intrinsic dimension estimate from the PCA eigenvalue spectrum.
Spectral decay	diagnostic	Slope of the sorted eigenvalue curve.
Participation ratio	higher	Effective number of active latent dimensions.
Anisotropy	diagnostic	Directional spread uniformity across the latent axes.
Noise resilience	higher	Embedding stability under Gaussian perturbation.
LSE overall	higher	Composite of normalized latent-space diagnostic scores.

Statistical testing

Two-sided Wilcoxon signed-rank tests compare paired method outputs across the same datasets.
Benjamini-Hochberg FDR correction is applied within each results table at q = 0.05.
All comparisons use the same preprocessing pipeline and a fixed random seed where subsampling is needed.

Continue

Metrics

Clustering

Projection fidelity

Latent structure

Metric inventory

Clustering quality

Dimensionality reduction evaluation

Latent space evaluation

Statistical testing

Benchmarks

Datasets

Code