Fig. 2

a Illustration of a typical pyHiM analysis on mouse tissues: examples of raw data are shown in the top row and the most relevant pyHiM outputs are shown in the bottom row. From left to right, raw DAPI data are segmented to compute the 3D masks of each individual nucleus. Next, 2D and 3D registration of the fiducial is performed for each imaging cycle, and the quality of the correction can be quickly assessed based on the output image. Then, the localization of individual DNA-FISH spots is performed in two steps: first, a 3D mask of each DNA-FISH spot is computed using deep learning. Then, using the mask position as a reference, the sub-pixel localization of the spot is inferred using apiFISH. Scale bars = 8 μm. b Chromatin tracks are calculated by combining all individual DNA-FISH spot localizations detected within the same mask (DAPI, or locus). Each individual trace represents a snapshot of the locus conformation within a single cell (see reconstruction with two different orientations). c Data quality assessment: (top) the N-matrix represents the number of times that each pair of DNA loci was detected in the dataset and is indicative of their detection efficiency. (bottom) The distribution of pairwise distances between DNA-FISH spots in the same chromatin trace is plotted to ensure that there is no major error in the analysis (detection threshold, etc.). d Traces computed by pyHiM were sorted based on RNA expression profiles in NC14 fly embryos and assigned to specific cell types (e.g., mesoderm vs. neuroectoderm). Specific long-range interactions and chromatin organization are observed for each cell type. e Fast 2D analysis based only on the projected 3D data is used to optimize parameters and test data quality. An example from mouse tissue data shows the pairwise distance maps computed using 2D (top) and 3D (bottom) analysis. The 2D map captures most of the features that characterize the conformation of the locus. f Comparison of pyHiM execution times for different number of cycles and for a desktop computer (Intel(R) Core(TM) i7-8700 CPU @ 3.20 GHz, CPUs: 12, cores: 6, threads per core: 2, memory: 16 Gb) or a multi-threaded server (AMD EPYC 7702 64-Core Processor 3.34 GHz, CPUs: 256, cores: 128, threads per core: 2, memory: 512 Gb). g Performance of pyHiM using single-threaded or DASK-powered multi-threading