Reversed graph embedding resolves complex single-cell developmental trajectories

Key results

1. ShownOn Paul et al. mouse myeloid progenitor data, Monocle 2 pseudotime agreed with a marker-based reference about as well as DPT and Wishbone and better than SLICER or Monocle 1, and its branch assignments stayed more accurate and consistent than Wishbone's under subsampling.

   Pearson's Rho, Kendall's Tau, adjusted Rand index (ARI) and lag-1 autocorrelation were computed on the same 300 sampled cells for all tools (SLICER and Monocle 1 could not run on all 2699); robustness used 25 subsamples of 80% of cells plus progressive downsampling, and outputs held across widely varied DDRTree parameters. No numeric scores are printed in the text. · Monocle 2 outperforms alternative reconstruction…; Comparing different algorithms to a marker-based…; Assessing robustness of pseudotime and branch…; Figure legends

2. SuggestedHuman skeletal myoblasts previously ordered as a linear trajectory resolved into one branch with two outcomes: F1 activates contractile genes and myogenic transcription factors, consistent with myotube fusion, whereas F2 does not, which the authors interpret as failed fusion.

   94 myoblast transcriptomes; BEAM found 664 genes differing between outcomes F1 and F2 (FDR < 10%), with TNNT2, MYH2, MEF2C and MYOG high on F1, while CCNB2 behaved similarly on both; Monocle 1 on the same genes gave a linear order. The unfused reading leans on earlier MYH2 immunofluorescence. · Unsupervised feature selection identifies key…; Monocle 2 accurately reconstructs single cell…; Analysis of HSMM data; Figure legends

3. ShownIn Olsson et al. mouse data, adding knockout cells kept a single major branch point but shifted cell fates: Gfi1-/- cells largely avoided the granulocyte branch, Irf8-/- cells the monocyte branch, and double knockouts sat on both branches near the branch point.

   A trajectory-conditioned test (knockout vs pseudotime-matched wild-type cells; FDR < 10%) found Gfi1-/- monocyte-branch cells over-expressing granulocyte genes and Irf8-/- granulocyte-branch cells over-expressing monocytic genes. In Paul et al. data, Cebpa-/- cells were absent from the erythroid branch and Cebpe-/- cells depleted from the GM branch. · Genetic perturbations divert cells along…; Trajectory-conditioned test; Figure legends

4. ShownOn Olsson et al. FACS-sorted mouse myelopoiesis cells, Monocle 2 learned one major branch: LSK cells at the tip chosen as root, CMPs on the trunk upstream of the branch, granulocytes and monocytes on separate branches, and mixed-lineage plus rare transient cells just upstream of the branch point.

   542 branch-dependent wild-type genes (BEAM, FDR < 10% per legend; the text states FDR < 1%); many were Irf8 or Gfi1 promoter targets by ChIP-seq, and both regulators bifurcated earlier in pseudotime than direct and motif-predicted secondary targets (FIMO on GMP DHS peaks). · Monocle 2 captures granulocytic/monocytic…; Analysis of blood data; Figure legends

5. SuggestedIn simulated cells from a 12-gene stochastic network producing neuron, astrocyte or oligodendrocyte fates, Monocle 2 recovered the three-fate tree, tracked the theoretical most-probable paths, and put branch points where Mash1/Hes5 and Scl/Olig2 turn mutually exclusive.

   Stochastic differential equations integrated by the Euler method for 400 time steps and run 200 times; one simulated path per fate was given to Monocle 2 at default settings (DDRTree in a two-dimensional latent space; SimplePPT in the original 12 dimensions) and compared graphically with ODE fixed-point paths; no error metric is printed. · Monocle 2 accurately reconstructs single cell…; Simulating the neuronal differentiation…; Figure legends