How Much Sequencing Is Enough? Study Maps the Depth Limits of Hi-C Chromatin Maps

Every cell in the body carries roughly two meters of DNA folded into a nucleus smaller than the width of a human hair, and the three-dimensional architecture of that fold is increasingly understood through a technique called Hi-C. The method captures millions of pairwise contacts between distant stretches of the genome and condenses them into genome-wide contact maps from which researchers extract everything from megabase-scale compartments to kilobase-scale chromatin loops. But a persistent question has haunted the field: how deeply do these experiments actually need to be sequenced before the numbers they produce mean anything? A new systematic analysis published in BMC Genomics by Alexander Granovsky and Kirill Polovnikov of the Skolkovo Institute of Science and Technology and the University of Potsdam now provides one of the most rigorous answers to date, and the results suggest that many labs may be spending sequencing budgets far less efficiently than they realize.

High-resolution Hi-C analyses have traditionally leaned on hundreds of millions of sequenced read pairs, a convention driven more by precedent than by evidence. The problem is that no one had systematically established the minimum depth required for the different classes of features that researchers routinely compute from contact maps. Granovsky and Polovnikov tackled this by performing progressive random subsampling of twelve Hi-C libraries spanning multiple cell types and experimental protocols. By repeatedly downsampling each library to different depths, they could track how each derived metric fluctuated as data became sparser, effectively simulating dozens of cheaper experiments from every expensive one already in hand.

The crucial innovation lies in how the authors defined sufficiency. Rather than declaring a depth adequate when a metric reached some arbitrary fraction of its full-depth value, they introduced a biologically motivated criterion grounded in the variability between independent biological replicates. The logic is elegant: if the scatter in a metric caused by reducing sequencing depth falls below the scatter observed between genuine replicate experiments, then adding more reads cannot improve the estimate any further, because biological noise has become the dominant source of variation. The required depth for each metric, then, is the point at which subsampling-induced scatter first reaches that replicate-level floor.

Applying this criterion across their panel of libraries produced strikingly different thresholds for different metrics. The genome-average loop-density estimate, derived from the log-derivative of the contact probability function P(s), reached its biological floor at approximately seven million read pairs. The genome-average loop-size estimate required roughly twenty million. Insulation scores, which quantify the depletion of contacts at boundaries between chromatin domains, needed around eighty million read pairs. And compartment eigenvectors, the principal descriptors of active versus inactive chromatin organization, demanded approximately one hundred ten million read pairs when computed at one-hundred-kilobase resolution. Compartment, insulation, and boundary analyses were performed on a subset of nine libraries with comparable full-depth coverage, while loop inference was evaluated across all twelve.

These numbers carry real weight for experimental design. Loop density and loop size, which reflect the fingerprint of the loop-extrusion process that shapes chromatin into loops anchored by CTCF and cohesin, turn out to be remarkably cheap to measure accurately at the genome-wide level. A lab interested primarily in global extrusion characteristics can obtain biologically saturated estimates with a small fraction of the sequencing typically deployed. In contrast, compartment eigenvectors at kilobase resolution sit at the expensive end of the spectrum, approaching the very depths that have become standard practice in high-resolution studies. The hierarchy of requirements means that a one-size-fits-all sequencing budget inevitably overpays for some measurements and may still underdeliver on others.

Perhaps the most consequential finding is that for every metric examined, depth-induced variability fell below biological replicate variability well before full sequencing depth was reached. In other words, the marginal value of additional reads on a single library diminishes to zero long before the deepest libraries stop yielding information. Once the subsampling scatter hits the replicate floor, extra sequencing refines nothing that matters, because the remaining uncertainty reflects genuine biological variation between cell populations rather than statistical noise. This reframes the entire question of depth: the ceiling is set not by the sequencer but by the biology itself.

The authors are careful to delineate the scope of their conclusions. The thresholds they report apply to genome-averaged characteristics, not to local metrics at specific genomic loci. A researcher hunting for a boundary shift at a particular disease-associated locus, or validating a single enhancer-promoter loop, faces a different statistical regime in which rare contacts must be sampled densely enough to resolve one narrow window of the genome. Genome-wide averages benefit from the law of large numbers across millions of loci, which is precisely why they saturate at comparatively modest depths, while locus-specific calls cannot borrow strength from the rest of the map in the same way.

The practical upshot for the field is a reallocation argument. If global biophysical features are the goal, the analysis suggests that increasing the number of biological replicates is likely to improve inference more effectively than sequencing individual libraries to greater depth. Multiple shallower libraries sample the true biological heterogeneity of a cell type, whereas one ultra-deep library samples a single biological state with precision that quickly becomes redundant. For groups planning Hi-C campaigns, this means the sequencing budget conversation should begin with the specific metrics to be computed, not with a blanket target of hundreds of millions of read pairs per sample.

The study also carries methodological significance for how the community benchmarks genomics pipelines. Subsampling-based calibration against replicate variability offers a general template that could be extended to other contact-map assays, such as Micro-C, and to other classes of derived quantities beyond those examined here. Because the work relied entirely on publicly available sequencing datasets, with no new human or animal data generated, it demonstrates how reanalysis of existing archives can yield design guidance of immediate value to the field. The research was supported by the Russian Science Foundation, with additional support to Polovnikov from the Alexander von Humboldt Foundation.

As Hi-C and its descendants become routine tools in chromatin biology, genome assembly, and disease genomics, the economics of sequencing decisions increasingly shape what questions laboratories can afford to ask. Granovsky and Polovnikov have converted a folk assumption about depth into a quantified, metric-by-metric benchmark, and in doing so they have handed experimentalists a map of where each additional million read pairs stops buying accuracy and starts buying only cost. For a field that has often equated more data with better science, the message is a bracing one: sometimes the deepest insight comes from knowing exactly when to stop sequencing.

Subject of Research: Sequencing depth requirements for Hi-C-derived chromatin metrics

Article Title: Systematic assessment of sequencing depth requirements for Hi-C-derived metrics

Article References: Systematic assessment of sequencing depth requirements for Hi-C-derived metrics. (n.d.). https://doi.org/10.1186/s12864-026-13455-7

Image Credits: AI Generated

DOI: 10.1186/s12864-026-13455-7

Keywords: Hi-C, sequencing depth, chromatin organization, A/B compartments, insulation score, loop extrusion, contact probability, biological replicates, benchmark, genomics, next-generation sequencing, chromatin structure