Fig. 2.

Fusing metrics of cell heterogeneity increases the percentage of validated connections. a When median, MAD and random projections of covariance profiles are combined through SNF (red line), the enrichment in having same MOA/pathway annotations is improved, especially for the strongest, most relevant connections above 0.5%. This is shown in three separate experiments involving small molecules (left, right) and gene overexpression (middle). Enrichment is versus a null distribution, which is based on the remainder of the connections. b Similarity graphs for the mechanism of action (MOA) class Adrenergic receptor antagonists, using different types of profiles in CDRPBIO-BBBC036-Bray. This MOA was chosen because it showed the highest improvement upon combining different profiles. The goal is a qualitative view on how data fusion improves within-MOA connectivities. Each node represents a compound, and two nodes are connected if the similarity of their corresponding profiles is ranked among the top 5% most-similar pairs. Median, MAD, and random projections of covariance profiles seem to be complementary for this MOA, as they cover mostly non-overlapping compound connections. The overall connectivity of compounds in this MOA is improved once these profiles are combined through SNF. Graph layouts are the same across data types and are based on the similarities in median + MAD + cov. (SNF); note that this causes the left-most graph to appear less cluttered and less connected, but the main purpose of the visualization is to observe the structure of connections, not the number of connections (which is quantified systematically in part a). c Weighted similarity graph as in the previous plot except that edge thicknesses are based on an exponential weighting of the ranked similarity values. Sub-clusters that are moderately present in two or three profile types (such as the one marked in red in bottom left) became stronger after applying data fusion using SNF