Fig. 5

Precision-Recall summary plots in Yeast and MCF7 breast cancer cell lines. These aggregate results were constructed by summarizing precision-recall confidence regions over every ss-DEG evaluation by reporting the best mean values with one standard deviation bars in each direction creating a cross, to create the broadest possible precision-recall combinations. The curves show a spectrum of operating characteristics across techniques, indicating the need for an ensemble-like approach and substantial improvements in ss-DEG. The MCF7 case study produced reference standards that predicted between 15% of the genes in the genome as DEGs, while the Yeast case study produced reference standards that predicted between 55 and 70% of the genes as DEG. The more clinically relevant range of DEGs from the MCF7 reference standard construction introduces a very distinct detection problem where methods like DEGseq result in a large number of False Positive as shown in the precision-recall summary plots. It achieves high recall at the expense of low-precision. Conservative techniques like DESeq obtain a very high precision on a small number of calls. The results show this is a challenging detection task, and that various techniques operate differently, providing an analyst with a wide-range of operating characteristics. In the Yeast dataset, all methods achieve a high precision, with varying levels of recall, however given that the majority of genes are labeled DEGs, this favors methods with high number of calls. Since certain methods can perform well in one scenario and underperform in others, we recommend a contextual use or an ensemble-like approach where the strengths of these tools can be combined into a single, robust predictor. Here, precision and recall of each instance of ss-DEGs are respectively calculated on the union and the intersection of reference standards (Table 2). Of note, at FDR < 20%, DESeq2 produces no predictions and is thus not shown and considered inappropriate for single-subject DEG analyses.