The PCA-based representative parameter selection currently projects each algorithm's reconstructed pathways onto an algorithm-specific PCA space and selects a representative parameter combination using only the first two principal components.
The two-component restriction is a choice in the PCA function code; however it is not a constraint of the KDE step. The KernelDensity model from scikit-learn fits data of any dimensionality, so it can run on more components. Specifically, our code computes PCA with pca_instance = PCA(n_components=components), where components can exceed 2, but it only fits the KDE only on xy = X_pca[:, :2]. As a result, the contour plot, peak search, and selected representative parameter combination use only PC1 and PC2 even when components > 2.
So something to consider is if we should representative selection using more than the top two PCs? Using additional components could capture more of the variance in pathway structure per algorithm to give us a higher-dimensional density estimate, but with maybe a loss of the 2D contour visualization.
The PCA-based representative parameter selection currently projects each algorithm's reconstructed pathways onto an algorithm-specific PCA space and selects a representative parameter combination using only the first two principal components.
The two-component restriction is a choice in the PCA function code; however it is not a constraint of the KDE step. The
KernelDensitymodel from scikit-learn fits data of any dimensionality, so it can run on more components. Specifically, our code computes PCA withpca_instance = PCA(n_components=components), wherecomponentscan exceed 2, but it only fits the KDE only onxy = X_pca[:, :2]. As a result, the contour plot, peak search, and selected representative parameter combination use only PC1 and PC2 even whencomponents > 2.So something to consider is if we should representative selection using more than the top two PCs? Using additional components could capture more of the variance in pathway structure per algorithm to give us a higher-dimensional density estimate, but with maybe a loss of the 2D contour visualization.