Bootstrapping developmental reconstruction
The MST process finds the path that connects all points in the dataset
with minimum total length, representing a putative developmental
trajectory through cellular ‘snapshot’ data. While this and other
graph-based strategies have been previously demonstrated to accurately
reconstruct uni-directional and branching developmental processes
(Bendall et al., 2014; Setty et al., 2016; Trapnell et al., 2014), the
presence of ‘short-circuits’, incorrectly drawn edges between cells in
different developmental stages, can cause significant errors in this
procedure. This concern is particularly relevant for MST construction,
which shares similarities with single-linkage clustering. To ensure that
our developmental reconstructions were not driven by these artifacts, we
performed the MST-building process on 1,000 subsamples of our data
(which each subsample containing 800 micro-clusters), and assessed the
reproducibility across bootstraps. We used the same reconstruction
procedure, consisting of MST construction based on diffusion map
coordinates followed by branch annotation, for each subsample.
When assessing our bootstraps, we found that in 1,000 subsamples, we
obtained identical hierarchical relationships as shown in Figure 2C.
Therefore, we conclude that the hierarchical relationships we derive
between HSC/MPP to the four downstream lineages are robust to potential
artifacts in the MST-building procedure.
However, our megakaryocyte (Mk) micro-clusters (Figure S2D) did not
exhibit consistent relationships across bootstraps. We observed that Mk
micro-clusters branched from different locations in the hierarchy in
different subsamples, resulting in multiple potential models for Mk
development (the relative proportion of subsamples leading to each model
is shown in Figure S2D). We therefore conclude that our dataset is
insufficient to resolve the precise location of Mk branching, and
excluded this lineage from further analysis. Therefore, the hierarchy we
propose in Figure 2C is consistent with the presence of a common
progenitor for Mk and other lineages, but also with the potential for Mk
to derive directly from HSC, as has been recently proposed (Grover et
al., 2016).