Choose more PC for the silhouette score, and run the deconvolution benchmark and see if it improves, out methods compared to cibersortX #14

Jianwu1

@stemangiola
Finalised benchmark: using new tree (treg as a sibling to t_helper) and the tip nodes for non_hierarchical and cibersortx:
PCA as reduction method;
Overall hierarchical_pairwise_contrast_bayes_silhouette_penalty gives cell signatures that result in the lowest median deconvolution error, it’s ranked first in dim=4 and dim=10, ranked third in dim=2, and the same method with curvature optimisation was ranked first in dim2.
Dim4 captures the most variance in the data while not compromising distance-based silhouette metric too much.
Methods that ranked higher than cibersortx are all hierarchical methods. This shows the effectiveness of hierarchical approach, however, another reason could be because more signatures are selected in hierarchical approach: 1126 genes selected in hierarchical_pairwise_contrast_bayes_silhouette_penalty_4, yet 286 genes selected in non_hierarchical_pairwise_contrast_bayes_naive_penalty_4 (didn’t run silhouette selection with non_hierarchical pairwise contrast methods due to the excessive computational load).

In terms of cell_type specific deconvolution error, cibersortx performs the best stably.
In terms of silhouette score, cibersortx performs the worst and non_hierarchical methods perform the best especically with silhouette selection since it’s the objective function.

Deconvolution
Dim=2: image Dim=4: image Dim=10: image

Silhouette:
Dim=2: image Dim=4: image Dim=10: image

Cell type:
Dim=2: image Dim=4: image Dim=10: image

Read more here: Source link