We used matrices describing each enhancer expression in TPMs for each facet (primary cell facets and tissue facets were clustered independently) and clustered these by complete linkage agglomerative hierarchical clustering using Euclidan distances, as implemented in the gputools R package62, and ran these in parallel on a GTX960 Nvidia GPU. Due to limited memory in the GPU, we reduced the matrices to enhancers with total expression > 2.5 TPM in the primary cell set and >0.6 TPM in the tissue/organ set, resulting in sets of roughly 22.500 enhancers each. To make sure these results were stable, we also explored normalization using fold change vs. background in each facet instead of TPM normalization, which resulted in very similar results (data not shown).
