Weibull Background Model Substantially Improves Chromatin Loop Detection in Genomic Maps

A computational study in Frontiers in Bioinformatics demonstrates that replacing the background statistical model in a prominent chromatin loop detection algorithm markedly expands the identification of three-dimensional genome structures. Developed by Benjamin Soibam of the University of Houston-Downtown, the modified computational workflow, designated WLoops, alters how local neighborhood variation is evaluated in high-throughput chromosome conformation capture contact matrices. Modern genomic research relies on Hi-C maps to detect chromatin loops, which form structural arcs physically connecting distant DNA segments. While the widely used software Mustache models background noise using an exponential distribution, Soibam’s evaluation establishes that a Weibull distribution provides a substantially superior fit to underlying image data. By integrating this statistical foundation into the existing multi-scale blob detection framework, WLoops uncovered thousands of previously undetected chromatin loops across human cell lines without altering candidate detection logic.

The physical organization of the eukaryotic genome within the nucleus is structured into dynamic loops regulated during diverse biological processes. Under the loop extrusion principle, these loops arise when distant genomic loci are brought into close proximity, isolating intervening sequences and forming functional architectural domains. These looping events serve foundational roles in cellular physiology, establishing enhancer-promoter contacts, enabling gene co-regulation, maintaining genomic insulation, and directing structural compactness. Because disruptions to these configurations are implicated in developmental abnormalities and disease, the precise genome-wide mapping of chromatin loops has become a central objective in computational genomics. Researchers analyze Hi-C contact maps, which represent interaction frequencies between discrete genomic bins, to identify these structural features in an unsupervised manner across diverse cellular contexts.

To detect chromatin loops computationally, bioinformatics pipelines often employ unsupervised local enrichment algorithms that identify statistically significant peaks within Hi-C contact maps without requiring prior labeled training data. Early tools such as HICCUPS, SIP, and cLoops evaluated contact matrices at fixed genomic representations, frequently mischaracterizing loops that varied in spatial scale. To address this limitation, the Mustache algorithm introduced multi-scale computer vision principles, treating Hi-C matrices as digital images and applying scale-space theory. Mustache detects loop candidates as blob-like structures by calculating local maxima in Difference-of-Gaussians filter responses across multiple scales. Unlike conventional loop callers that evaluate significance directly from raw counts, Mustache assesses candidate significance from Difference-of-Gaussians values, treating loops strictly as image features. However, to distinguish biological signals from background variation, Mustache modeled absolute Difference-of-Gaussians values using an exponential distribution without empirical justification.

To determine whether local background noise actually conforms to an exponential curve, Soibam conducted an empirical distribution fitting analysis across five-kilobase resolution Hi-C maps from human GM12878 lymphoblastoid cells and K562 erythroleukemia cells. Using the distfit Python package, four candidate models were evaluated across each local contact block: exponential, Weibull, Pareto, and log-normal distributions. Parameters were estimated by minimizing the residual sum of squares between theoretical curves and empirical Difference-of-Gaussians absolute values. The Weibull distribution consistently yielded the lowest average residual sum of squares in both cell lines, followed by the exponential distribution. Evaluating model complexity using the Akaike information criterion, which penalizes parameter proliferation, showed that the Weibull model achieved lower scores. Furthermore, the Anderson-Darling statistic, which weights distribution upper tails where significance thresholds operate, revealed substantially poorer fit values for the exponential model, confirming the Weibull distribution’s superior accuracy.

Having demonstrated the statistical superiority of the Weibull distribution, the study examined whether swapping the background model altered the quantity and identity of detected chromatin loops. Using Mustache’s original blob detection architecture, loop calling was performed on the GM12878 and K562 five-kilobase datasets under both background distributions, retaining candidate blobs with a false discovery rate below zero point one. In the GM12878 cell line, the exponential background yielded 16,350 loops, whereas the Weibull background identified 23,159 loops. Among these, 15,754 loops were shared between distributions within a five-kilobase anchor matching window, while 7,405 loops were uniquely discovered by the Weibull model. In the K562 cell line, the Weibull model detected 12,859 loops compared to 10,089 identified by the exponential distribution, uncovering 3,321 unique loops while recovering 9,538 shared interactions. Notably, none of the Weibull-unique loops overlapped with benchmark detections published using HICCUPS.

To evaluate whether newly identified loops represented authentic physical interactions rather than statistical artifacts, the study performed aggregate peak analysis across both loop populations. Aggregate peak analysis quantifies contact enrichment by superimposing twenty-one by twenty-one pixel submatrices centered on detected loop anchors, restricted to interactions spanning 150 kilobases to one megabase. In aggregate peak analysis heatmaps, high-confidence chromatin loops produce a pronounced central peak relative to surrounding pixels, quantified as an aggregate peak analysis score dividing the center pixel value by surrounding pixel intensities. Across both GM12878 and K562 datasets, exponential and Weibull loop sets exhibited robust central enrichment and scores comparable to published benchmarks. However, when evaluating loops uniquely detected by each model, the Weibull-specific loops demonstrated distinctly sharper, more focused central enrichment and slightly higher aggregate peak analysis scores than exponential-unique loops, which showed more diffuse enrichment.

The study further evaluated whether the novel loops possessed biological characteristics indicative of authentic chromosome architecture by examining their spatial correspondence with CCCTC-binding factor, or CTCF. Chromatin loops are predominantly organized and anchored through physical binding of CTCF alongside the structural Cohesin complex. Analyzing published CTCF chromatin immunoprecipitation sequencing data from the ENCODE repository, the author extended detected loop anchors by fifteen kilobases and examined physical binding overlap. In GM12878 cells, 92 percent of Weibull-detected loops and 96 percent of exponential loops overlapped CTCF binding peaks, with Weibull-unique loops displaying an 88 percent overlap rate. In K562 cells, Weibull and exponential loops showed 85 percent and 86 percent CTCF overlap, respectively, with Weibull-unique loops maintaining an 85 percent overlap. Furthermore, scanning anchors for directional CTCF motifs using the FIMO software against the JASPAR database revealed convergent motif pairs in 77 percent of shared GM12878 loops and 63 percent of Weibull-unique loops.

In addition to CTCF profiling, the investigation assessed whether detected loops coincided with the structural Cohesin complex, which collaborates directly with CTCF to stabilize extruded chromatin loops. The author integrated public chromatin immunoprecipitation sequencing datasets for RAD21 and SMC3, two essential core subunits of the human Cohesin ring. In the GM12878 cell line, 77 percent of loops identified by the Weibull distribution overlapped combined RAD21 and SMC3 binding sites, compared to 88 percent of loops called using the exponential background model. When isolating loops detected exclusively by a single model, 67 percent of Weibull-unique loops and 80 percent of exponential-unique loops exhibited Cohesin subunit occupancy. In the K562 cell line, the correspondence remained consistent, with 74 percent of Weibull-unique loops demonstrating Cohesin binding at anchor loci, compared to 87 percent for exponential-unique loops, confirming structural protein occupancy at anchor sites.

To corroborate the Hi-C findings against independent experimental technologies, the study compared detected loop sets to orthogonal molecular datasets that enrich specifically for protein-mediated contacts. Using GM12878 data, the author cross-referenced loop coordinates against chromatin interaction analysis by paired-end tag sequencing, or ChIA-PET, targeted to CTCF, alongside HiChIP interaction maps targeted to Cohesin. Approximately 34 percent of the loops unique to the Weibull background model aligned directly with CTCF ChIA-PET interactions, while 24 percent matched Cohesin HiChIP interactions. In comparison, loops unique to the exponential distribution matched 30 percent of CTCF ChIA-PET and 42 percent of Cohesin HiChIP interactions. Although the Cohesin HiChIP concordance was lower for Weibull-unique interactions than for exponential-unique ones, the author noted that these recovery proportions are fully congruent with concordance rates reported by alternative published loop callers, such as findings documented by Chowdhury and colleagues.

The author subsequently assessed the generalizability of the Weibull background model across distinct experimental protocols, additional mammalian cell lines, and varying contact matrix resolutions. Applying WLoops to Hi-C data from human umbilical vein endothelial cells and micro-C data from mouse embryonic stem cells, the Weibull model successfully recovered 93 percent and 96 percent of exponentially called loops, respectively. Concurrently, it identified 1,219 novel unique loops in endothelial cells and 1,156 in mouse embryonic stem cells, representing a two- to threefold increase in unique loop discovery compared to the exponential distribution. Both additional datasets demonstrated higher aggregate peak analysis scores and sharper central enrichment for Weibull-specific loops. Furthermore, re-evaluating GM12878 and K562 Hi-C maps at a coarser ten-kilobase resolution recovered 94 to 95 percent of exponential loops while detecting 1,175 and 1,864 unique loops, respectively, with persistent sharp aggregate peak analysis enrichment.

Addressing theoretical considerations, the paper highlights that the two-parameter Weibull distribution provides mathematical flexibility that the single-parameter exponential distribution lacks. The Weibull distribution features a variable shape parameter that accommodates non-rigid failure rate behavior and reduces to an exponential form when the shape parameter equals one. With sufficient genomic contact resolution, this flexibility models local background variation accurately without inducing overfitting. While selecting optimal distribution models independently for every individual sub-region across a contact matrix could theoretically optimize detection further, such an approach would dramatically increase computational time. The paper cautions that loops unique to the exponential distribution should not be dismissed entirely, as they retain biological support despite lower contact enrichment. Instead, the author concludes that WLoops serves as an essential tool to supplement and refine existing detection pipelines, releasing the modified code publicly on GitHub to facilitate enhanced genomic mapping.

Subject of Research: Computational detection of chromatin loops from Hi-C contact maps using statistical background modeling

Article Title: WLoops: appropriate background model in scale-space representation of Hi-C maps improves detection of chromatin loops

Article References: Soibam, B. (2026). WLoops: appropriate background model in scale-space representation of Hi-C maps improves detection of chromatin loops. Frontiers in Bioinformatics, 6, 1986674. https://doi.org/10.3389/fbinf.2026.1986674

Image Credits: AI Generated

DOI: 10.3389/fbinf.2026.1986674

Keywords: Chromatin Loops, Hi-C Maps, Weibull Distribution, WLoops, Mustache Algorithm, Computational Genomics, CTCF Binding, Cohesin Complex, appropriate, background, model, scale-space