Fig. 6. Restriction augments biomarker discovery in synthetic and real-world flow cytometry datasets.

Using synthetic and real-world datasets, we demonstrate that restriction augments the discovery of biomarkers with limited informative ranges. In each plot, the x-axis shows permutation p-values for the AUC of complete ROC curves for every gated cell population. The y-axis shows permutation p-values for the AUC of optimally restricted ROC curves. Points within the green-shaded rectangles represent cell subsets whose p-values derived from unrestricted data are not significant (p ≥ 0.05), but p-values derived from optimally restricted data are significant (p < 0.05). a Permutation p-values from synthetic samples in which a disease-related effect was introduced into CD4⁺ T_EM resulting in a symmetric ROC curve. 100 samples in the negative $\mathcal{N}\left(7.7,1\right)$ class and 100 samples in the positive $\mathcal{N}\left(10.7,1\right)$ class were generated. b Permutation p-values from synthetic samples in which a disease-related effect was introduced into CD4⁺ T_EM resulting in a right-skewed ROC curve. 100 samples in the negative $\mathcal{N}\left(7.7,1\right)$ class and 100 samples in the positive $\mathcal{N}\left(8\mathit{.}7,3\right)$ class were generated. c Permutation p-values from synthetic samples in which a disease-related effect was introduced into CD4⁺ T_EM resulting in a left-skewed ROC curve. 100 samples in the negative $\mathcal{N}\left(7.7,3\right)$ class and 100 samples in the positive $\mathcal{N}\left(8.7,1\right)$ class were generated. d Permutation p-values from synthetic samples in which a disease-related effect was introduced into CD4⁺ T_EM resulting in a right-skewed ROC curve. 100 samples in the negative $\mathcal{N}\left(7.7,1\right)$ class and 100 samples from a bimodally distributed positive class were generated. In this example, the positive population comprises 20% cases with elevated biomarker expression $\mathcal{N}\left(16.7,1\right)$ and 80% cases with unaltered biomarker expression $\mathcal{N}\left(7.7,1\right)$. e Permutation p-values from a training set of real-world clinical flow cytometry samples (n = 110). 84 biomarkers were selected where ≥ 10% of samples had more than 10 counts. Dataset restriction reveals 4 previously undescribed biomarkers of treatment-related colitis risk in metastatic melanoma patients receiving Ipi-Nivo therapy. f Permutation p-values from a training set of real-world clinical flow cytometry samples (n = 110). 84 biomarkers were selected where ≥ 10% of samples had more than 10 counts. Dataset restriction reveals 7 previously undescribed biomarkers of treatment-related hepatitis risk in metastatic melanoma patients receiving Ipi-Nivo therapy.