Table 1 Results of unsupervised dimension reduction techniques on the real dataset.

Method	Paroni Sterbini	Trust	Amir3	Trust	Amir4	Trust	mean
PSI-ROC
HD	0.88	0.0036	0.95	0.0009	0.98	0.0009	0.94
MDSwUF	0.84	0.0089	1.00	0.0009	0.88	0.0329	0.90
MCE	0.91	0.0036	0.88	0.0329	0.91	0.0009	0.90
PCA	0.85	0.0063	0.91	0.0009	0.86	0.0169	0.87
MDStyc	0.84	0.0076	0.88	0.0009	0.84	0.0249	0.85
NMDS	0.85	0.0036	0.86	0.0169	0.84	0.0089	0.85
MDSbc	0.81	0.0183	0.86	0.0089	0.84	0.0189	0.84
PSI-PR
HD	0.94	0.0009	0.96	0.0009	0.99	0.0009	0.96
MDSwUF	0.88	0.0036	1.00	0.0009	0.90	0.0089	0.93
MCE	0.96	0.0009	0.89	0.0089	0.92	0.0039	0.92
PCA	0.91	0.0039	0.90	0.0009	0.88	0.0089	0.90
MDStyc	0.88	0.0116	0.90	0.0009	0.88	0.0089	0.89
MDSbc	0.86	0.0116	0.89	0.0009	0.90	0.0009	0.88
NMDS	0.90	0.0036	0.87	0.0089	0.87	0.0009	0.88

Best results of unsupervised dimension reduction techniques according to the PSI indices for sample separation in the space of the first two dimensions of embedding. HD (no dimension reduction) represents the reference results to see how good the separability present in the high-dimensional space is preserved by dimension reduction techniques. Results are ordered from the best (top) to the worst (bottom) method. For the Paroni Sterbini dataset, we show the results for three different labels (PPI treated, untreated H+ and untreated H−). For the Amir datasets, the PSI measures were computed for two groups, identified by the presence or absence of PPI treatment. For each PSI value, a respective trustworthiness was calculated.
All PSI-ROC and PSI-PR values can be found in Supplementary Data 2.
HD high dimension, MCE Minimum Curvilinear Embedding, MDSbc Multidimensional Scaling with Bray-Curtis dissimilarity, MDSwUF Multidimensional Scaling with weighted UniFrac distance, NMDS Nonmetric Multidimensional Scaling, MDStyc Multidimensional Scaling with Theta-YC distance, PCA Principal Component Analysis, PSI-ROC Projection Separability Index measured by Area Under the Curve, PSI-PR Projection Separability Index measured by Area Under the Precision Recall, Trust trustworthiness.

Search