While our Drop-seq experiments focus on RNA expression, epigenetic changes, such as the remodeling in chromatin accessibility, are primary determinants of cellular potential. Given the conservation of trajectories between bone marrow and cord blood, we believed that cell fate transitions could be further investigated using open chromatin data collected from human bone marrow. We therefore leveraged a recently published ATAC-seq dataset from human hematopoiesis (Corces et al., 2016) to integrate chromatin dynamics alongside our transcriptional models. Chromatin accessibilities from HSC, MPP, LMPP, CLP, GMP, and MEP were used for our analyses, as we have shown (Figure 1C) that our Drop-seq clusters are transcriptomically similar to traditional gating for these cell types. Though we did not discover a ‘CMP’ cluster, we included this population in our ATAC-seq analyses as well, as we expect this gate to comprise a heterogeneous mix of EMP and erythroid-committed cells.
We first asked whether the hematopoietic hierarchy calculated from our transcriptomic data was reflected in chromatin accessibility. Each open chromatin region was annotated and linked to the closest transcription start site (TSS), and we identified the 2,000 most variable regions as inputs for principal components analysis (PCA) across progenitor types, which demonstrated that PCs 1 and 2 echoed the structure of our predicted hematopoietic hierarchy (Figures 4F?). Consistent with this, when we projected ATAC-seq peaks that were adjacent to genes from our transcriptomic clusters onto this PCA, we found that EMP-dependent genes and LMPP-dependent genes projected to opposing sides of the PC1-axis (Figure 4G). Notably, while the ATAC-seq dataset did not contain specific sorting of Ba/Eo/Ma progenitors, genes associated with this lineage showed similar PC1 scores to genes involved in erythroid differentiation, while Neu/Mo-associated genes grouped with lymphoid regulators. Therefore, hematopoietic fate transition is reflected on global chromatin remodeling that is consistent the transcriptomic changes identified from Drop-seq data.
We next asked whether peaks adjacent to ‘primed’ and ‘de novo’ genes exhibited distinct dynamics at the chromatin level. To integrate the accessibility data with our transcriptomic signatures, we first summarized the accessibility of each gene based on a single adjacent peak with the largest maximum accessibility across all cell types (‘primary peak’; STAR Methods). We then explored how the average accessibility of ‘primary peaks’ for each gene module changed across progenitor states (Figures 4C and S4B). ATAC-seq peaks for ‘primed’ genes exhibited high levels of accessibility in early progenitors, with HSC and LMPP exhibiting multi-lineage epigenetic priming for these loci as well. Accessibility was maintained, however, only during the transition towards a single lineage; for example we observed sharp decreases in accessibility for primed lymphoid genes in the LMPP to GMP transition, though accessibility was maintained in CLP progenitors. Genes in ‘de-novo’ modules exhibited low accessibility in upstream progenitors, and were specifically remodeled upon transcriptional activation (Figure S4B). Therefore, early decisions represent a coordinated and self-reinforcing process, where global epigenetic remodeling and transcriptional changes generally work in concert to direct and restrict cellular fate.