Calling structures from single cells¶

A single cell’s contact map holds 10⁴–10⁶ contacts — far too few for the callers of bulk Hi-C on that cell alone. Structures are still within reach, by four routes; which one works depends on the structure and on how many cells of a kind there are:

Route	What it calls	When	API
pool the cells, call the map	compartments, domains, loops, pileups	cell groups (types, clusters) with enough contacts: compartments and domains from ~2 % of bulk depth, loops near bulk depth	`uc.tl.pseudobulk` → `uc.tl.call_compartments(method="eig")`, `call_tads(method="insulation")`, `call_loops(method="hiccups")`, `uc.tl.pileup` (pseudo-bulk maps, calling structures from a map)
across cells, without pooling	loops (SnapHiC); domain boundaries per cell and their consensus	≥ ~100 cells of a type at 10 kb: SnapHiC finds loops where the pooled map gives almost none	`uc.tl.call_loops(cd, method="snaphic")`, `uc.tl.single_cell_boundaries(cd)`
per cell	compartment values (scA/B) of every cell	always; they separate cell types even at 10⁴–10⁵ contacts	`uc.tl.single_cell_compartments(cd, track=<genome FASTA>)`
on reconstructed 3-D structures	the tracing callers (compartments, TADs, loops on distances)	compartments: yes; TADs and loops: not at tens of cells	`uc.tl.reconstruct_sc` / published structures → `uc.tl.call_compartments(method="axes_pc")`

import uchrom as uc
import uchrom.datasets as ds

cd = ds.load("nagano2017_dip_serum")                    # 1,175 diploid mES cells, 10 kb maps per cell (linked)
top = cd.cells.index[cd.cells["snaphic_top742"]]         # SnapHiC's selection: the 742 deepest cells
loops = uc.tl.call_loops(cd, method="snaphic", cells=top)          # loops across the cells
bounds = uc.tl.single_cell_boundaries(cd, cells=top)               # boundaries per cell + consensus
scab = uc.tl.single_cell_compartments(cd, track=ds.fetch("mm10_genome"))   # cells x bins

Tutorial¶

Calling structures from single cells

Loops across cells: SnapHiC¶

Each cell’s map is imputed by a random walk with restart on its binary contact graph (whole chromosomes; or SnapHiC2’s sliding windows, ImputeParams(window="auto")), turned into per-diagonal z-scores, and every candidate pixel is tested against its local background over the cells (paired t-test, FDR per distance, the five filters and the clustering of Yu et al. 2021). The imputation, z-scores and per-cell sums run in Rust (uchrom-maps); without it the numpy code runs, which cannot hold whole chromosomes in memory — use window="auto" then.

Checked on the paper’s data (Nagano 2017 mES cells, benchmarks/sc_structures/snaphic_*.py):

the authors’ code: on the same 100 cells and contacts U-Chrom calls exactly their 479 loops (with their sliding-window quirk switched on; z-scores within 1.4e-13) — in 5 s instead of 21 min;
742 cells, genome-wide, vs the paper’s reference loops (HiCCUPS on bulk Hi-C + MAPS): recall 0.60 (paper 0.60), precision 0.60 (paper 0.66), 19,324 loops (paper 15,896); 73 % of the paper’s loops within 20 kb of ours. HiCCUPS on the pooled map of the same cells: 396 loops, recall 0.06.

Domain boundaries of single cells¶

The insulation score of each cell’s imputed map gives that cell’s boundaries; the cells’ mean insulation gives consensus boundaries. 742 mES cells vs bulk insulation: 60 % of the consensus boundaries within 2 bins of a bulk boundary (chance 16 %), 58 % of the bulk boundaries recovered; a single cell’s boundaries are rarely at the bulk ones (frequency 0.058 vs 0.040 elsewhere) — domains vary from cell to cell.

Compartments of single cells: scA/B¶

Per cell and bin, the mean CpG frequency of the bin’s contact partners (Tan et al. 2021, Dip-C; ranked per cell; ScABParams(contacts="cis", normalize="none") gives scHiCluster’s score). Identical to Dip-C’s and scHiCluster’s own code on the same contacts; 32,777 dscHi-C cells in 70 s. It separates cell types — Tan 2021 cortex structure types kNN 0.875 (scHiCluster embedding 0.68), dscHi-C cell types 0.91 — and follows bulk E1 (GM12878 cells vs Rao 2014: r 0.82 per cell). Much of it is the CpG track itself (its mean correlates 0.84–0.89 with CpG): the cell-type-specific part is real but modest (partial r given CpG 0.3–0.67).

The tracing callers on reconstructed structures¶

A reconstruction is a ChromData with coordinates — the tracing callers run on it unchanged (io.read_3dg_cells reads Dip-C / hickit structures; ds.load("tan2018_gm12878") holds the authors’ Dip-C structures). On 14 GM12878 cells (28 chr21 homologs, 20 kb) against Rao 2014:

Structure	Structures (tracing callers)	Same cells pooled (map callers)
compartments	r 0.77 with E1, agreement 0.86	r 0.93, agreement 0.95
domain boundaries	ArcFISH at chance; FISHnet frequency peaks 0.34 (chance 0.12)	insulation: precision 0.38, recall 0.70
loops	none called (axis-wise F), although published loop anchors sit at 0.69× the background distance	HiCCUPS: 4

Compartments are limited by the number of cells (the same as imaging at equal numbers of traces); domains and loops by the structures — a reconstruction is smoother than an imaged trace and the tracing callers’ tests treat it as data. For domains and loops pool the cells or call across them (above).

The numbers, datasets and Sherlock jobs are in benchmarks/sc_structures/README.md.