Spliced and Unspliced mRNA Quantification

To obtain separated count matrices for spliced and unspliced mRNAs, we ran velocyto 0.17.17 [10 (link)] on the .bam files from the mouse atlas in Pijuan-Sala et al. ([25 (link)]; Arrayexpress accession number: E-MTAB-6967). We kept all cells that passed the QC as described in the original publication, but filtered out from downstream analysis the extraembryonic tissues: ExE endoderm, ExE ectoderm, and Parietal endoderm as well as samples with no timepoint allocation (labelled as “mixed gastrulation”). To select highly variable genes (HVGs) we applied both the scanpy v1.5.1 and the scVelo v0.2.1 [14 ] pipelines. That is, we removed genes with less than 20 shared counts between spliced and unspliced counts, before normalizing and log transforming the remaining genes. Then, we selected the top 2500 HVGs from each approach (resulting in a total of 4000, with 1000 overlapping genes) for further calculation of moments, while performing imputation using the top 30 nearest neighbours from the graph connectivities generated with the original UMAP coordinates from Pijuan-Sala et al. [25 (link)]. The velocity vectors were computed in dynamical mode rather than steady state.