Fig. 3: Simulation study results.
From: Trajectory-based differential expression analysis for single-cell sequencing data
PCA plots for the (a) cyclic, (b) bifurcating, and (c) multifurcating simulated trajectories. The plotting symbol for each cell is colored according to its true pseudotime; trajectories (in black) were inferred by princurve in (a) and slingshot in (b) and (c). d–f Scatterplot of the true positive rate (TPR) vs. the false discovery rate (FDR) or false discovery proportion (FDP) for various DE methods applied to the simulated data sets. Panel (d) displays the average performance curves of DE methods across seven out of ten cyclic data sets for which all DE methods worked (Monocle 3 errored on three data sets). The associationTest from tradeSeq has superior performance for discovering genes whose expression is associated with pseudotime, as compared with Monocle 3. When investigating differential expression between lineages of a trajectory, the patternTest of tradeSeq consistently outperforms the diffEndTest across all three TI methods, since it is capable of comparing expression across entire lineages. Panel (e) displays the average performance curves across the three bifurcating data sets where all TI methods recovered the correct topology. Here, all tradeSeq patternTest workflows, tradeSeq_slingshot_end, and edgeR have similar performance and all are superior to BEAM, ImpulseDE2, and GPfates. Note that the performance of tradeSeq_Monocle2_end deteriorates as compared with tradeSeq_slingshot_end; the curve for tradeSeq_GPfates_end is not visible in this panel due to its low performance. For the multifurcating data set of panel (f), tradeSeq_slingshot has the highest performance, closely followed by tradeSeq_Monocle2 and edgeR.
