Can Coarse Contact Maps Locate a Ten-Base-Pair Hinge?

When DNA is twisted hard enough, it folds back on itself into a looped coil called a plectoneme, much like an over-twisted phone cord. At the sharp turn at the tip of that coil, a few base pairs can briefly come apart. The resulting single-stranded "bubble" acts as a floppy hinge. The DNA then no longer has to bend its stiff double helix around the turn, and the coil stops sliding freely along the molecule: it gets stuck at sequences that melt easily. The hinge itself spans only tens of base pairs. The coil it anchors, however, spans a kilobase or more, and it leaves a distinctive stripe in a chromosome contact map, the pairwise "how often are these two positions touching" matrix measured by Hi-C sequencing.

This note asks a concrete question: at what bin width and sequencing depth can a hinge's position be recovered from binned, sparse contact counts? And do the deep-learning "super-resolution" tools now used to sharpen such maps preserve this physical signal, erase it, or invent it? The key observation is that pinning concentrates many cells' conformations onto one sequence-determined position. That concentration should make sub-bin localization possible, in the same way a fluorescence microscope can localize a molecule far more finely than its pixel size.

Core Claim

The claim is that a hinge-pinned plectoneme is identifiable from contact data whose bins are much wider than the hinge, as long as the coil's footprint exceeds the bin width. It should be recoverable by a physics-based likelihood that models three things: sequence-dependent local melting, hard-core self-avoidance, and binning plus read sampling. A homogeneous polymer, a sequence-composition heuristic, or an image-trained super-resolution network should not recover it as well.

If the hinge is a localized target whose signature is spread over a footprint of many bins, the localization error $\sigma_h$ should scale roughly as

$$\sigma_h \approx \frac{c\, w}{\sqrt{N\, \Delta}},$$

where $w$ is the bin width, $N$ is the number of read pairs falling in the coil's footprint, $\Delta$ is the fractional contrast between pinned and unpinned stripe profiles, and $c$ is an order-one constant. This form comes from the standard centroid-localization argument: the error of a position estimate falls as one over the square root of the effective count. It breaks down when $w$ approaches the footprint size, and finding that breakdown point is the goal of the study.

The contribution is therefore an identifiability map plus an estimator. It is not another super-resolution network. The prior work cited in the file does not provide either piece. Plectoneme tip bubbles showed with the nucleotide-level oxDNA model that tip bubbles pin plectonemes 34–100% of the time and prefer AT-rich blocks. It worked only on 600 bp molecules and never asked what a contact map would show. Hi-C super-resolution methods such as DFHiC, HiCSR, HiCMamba and Capricorn optimize pixel or loop metrics with no physical ground truth. As a result, nobody can say whether they preserve mechanically real features.

Why This Combination, and What Each Direction Contributes

The transient-hinge literature supplies the mechanism. Beyond the tip-bubble study, CSO-EPBD cyclization modeling and the J-factor analysis in Evaluating the role of coherent delocalized phonon-like modes show that local base opening explains cyclization rates orders of magnitude above what a uniform elastic rod predicts. This is evidence that a homogeneous-stiffness polymer is the wrong null model at short scales.

The excluded-volume literature is a constraint rather than a decoration. Plectonemes are exactly the regime in which phantom rods pass through themselves; Bouchiat's continuous model fails there, and the discrete rod-like chain fixes this with hard cylinders. The discrete chain, however, keeps a uniform persistence length and has no hinge. Self-avoidance also matters for the contact statistics themselves. Polovnikov, Nechaev and Tamm derive a roughly 7% boost in triple-contact probability for Gaussian fractal chains with $d_f=3$, and they note explicitly that hard-core exclusion should push the other way. A hinge-bearing, self-avoiding plectoneme brings three arms together at its tip, so it is a natural local test of which effect wins.

Genomic granularity is the measurement regime. Here BRACIL (Decoding ChIP-seq with a double-binding signal) supplies the methodological template: an explicit physical model of how sequencing blurs a signal, inverted to reach single-nucleotide resolution in one dimension. This proposal does the analogous thing in two dimensions, with a polymer ensemble in place of a shearing distribution. Bayesian ISD ensembles supply the inference style, sampling ensembles with a likelihood rather than fitting one structure. That work uses 15 kb beads, and we move to roughly 10 bp segments.

Research Program

The first component is a mesoscale hinge-aware chain. Each segment covers about 10 bp and carries a two-state variable, paired or melted. Melting has a sequence-dependent free-energy cost and lowers the local bending and twist stiffness toward single-stranded values. Non-adjacent segments interact through hard cylinders, as in the discrete rod-like chain. We calibrate the melting and stiffness parameters by reproducing the oxDNA observables on the same 600 bp block-random construct (120 bp blocks at 70% and 30% GC): the pinning fraction as a function of tension, the preference for AT blocks, and the extension shift. Only after this calibration do we scale to 10–50 kb segments under negative superhelical density. That size is out of oxDNA's reach but tractable at this coarser level. A central risk is that the mesoscale model fails to reproduce oxDNA pinning. We check this before anything else.

The second component is a synthetic contact benchmark with known ground truth. Sampled ensembles give contact maps at 10 bp resolution, together with the true hinge-position distribution and three-body contact frequencies. A forward model then bins these at 100 bp, 1 kb, and 5 kb, and draws read pairs at depths matched to the full, 1/16, 1/32, and 1/64 downsampling levels used on GSE63525-style evaluations. Each sequence design has a matched control: the same sequence with melting disabled, and the same sequence with self-avoidance disabled. This lets us attribute any signal to one mechanism.

The third component is the estimator. We infer hinge position and pinning strength by maximizing, or sampling, the likelihood of binned counts under the forward model. To keep this tractable, we precompute ensembles on a grid of candidate hinge positions and interpolate. The output is a posterior with a credible interval, which the image models do not provide.

Baselines

The naive reference point is a sequence-composition heuristic. It predicts the hinge at the lowest-GC 100 bp window inside the observed stripe, and it predicts the stripe from nothing but the binned counts. If this matches the likelihood estimator, the physics adds nothing beyond "AT-rich melts first." Beating it requires lower localization error in designs where the weakest window is not where pinning actually occurs, for example where tension or the geometry of neighboring blocks shifts the optimum. The oxDNA results make such designs plausible.

The strongest existing method is image super-resolution followed by peak finding. We will use HiCMamba, whose reported CTCF-supported loop match rate is 66.4% versus 58.0% for HiCARN (HiCMamba), and Capricorn, which reports 25.6% higher cross-cell-line loop F1 on average (Capricorn). There is no reported result on Lacuna for hinge or plectoneme localization by any of these tools, so we will establish that baseline ourselves. For fairness, we will run each network twice. The first run uses released weights trained on GSE63525, to measure what current practice does to physical features. The second run retrains or fine-tunes the network on our synthetic training split, with the same training data and a tuning budget in GPU-hours matched to the cost of our simulation plus inference. It then predicts hinge position from the output map with the same peak-finder the other methods use. Beating it means lower localization error at equal depth and bin width, calibrated intervals, and fewer false stripes on melting-disabled controls.

The key-ingredient ablation is the same likelihood estimator run with a homogeneous self-avoiding rod: identical inference, but no melting variable. A second ablation removes hard-core exclusion while keeping the hinges. These two isolate whether the hinge, the self-avoidance, or both carry the identifiability.

Why not just…?

Why not just train a super-resolution network on the simulated maps and skip the likelihood? This is the real threat. A network trained on our simulations could localize hinges as well as the estimator on in-distribution data. We will report that honestly if it happens. The proposal's central output, the identifiability map, does not depend on which estimator wins, because it uses the best estimator at each operating point. What the likelihood adds is calibrated uncertainty, plus the ability to test at modest cost whether the answer holds under changed salt, tension, or sequence designs that were never in training. We will measure both methods on held-out designs with shifted superhelical density, where we expect the trained network to degrade faster. If it does not, we will recommend the network.

Why not just use a homogeneous self-avoiding polymer, as in the discrete rod-like chain? Bouchiat's model handles buckling correctly, but by construction it lets plectonemes slide. The tip-bubble study attributes pinning specifically to melting. The homogeneous-rod ablation directly tests whether a uniform chain produces sharp, sequence-anchored stripes; we predict it produces blurred ones.

Why not just correlate GC content with contact enrichment on real Hi-C? In real nuclei this conflates nucleosome positioning, transcription and protein binding with mechanics. Even on synthetic data, it cannot separate the melting and geometry effects described above. It remains the naive baseline, and it will win wherever the weakest window and the pinning site coincide. The benchmark reports those cases separately.

Proposed Experiments

The first decisive check is calibration. The mesoscale chain must reproduce oxDNA pinning within the 34–100% range across the same tensions, and it must show AT-block preference on the 600 bp construct. If it does not, the scale-up is unfounded.

The second check is the identifiability map. We plot hinge localization error against bin width and depth for all methods and test whether the likelihood estimator follows the predicted $w/\sqrt{N}$ trend. We also locate the bin width at which localization fails. We expect that point to be near the plectoneme footprint, roughly the 1–5 kb range where standard Hi-C operates, which is why the answer matters.

The third check is a hallucination audit. We apply released HiCSR, HiCMamba and Capricorn weights to melting-disabled and self-avoidance-disabled controls, then measure false stripe calls and shifts in stripe position relative to ground truth.

The fourth check is a local test of the multi-contact tension. On simulated ensembles, we measure the ratio of triple-contact probability to the product of pairwise probabilities near plectoneme tips. We do this with and without hard-core exclusion and with and without hinges, and compare against the roughly 7% Gaussian boost from Polovnikov et al.. A suppression below one with exclusion on would give a concrete, local resolution of the sign question they left open.

A final exploratory step asks whether the estimator's posterior over pinning sites shows any enrichment near actively transcribed, supercoiled regions in GSE63525 and the Micro-C data used by YOLOOP. It is framed as hypothesis-generating only, because nucleosomes and ionic conditions in vivo lie outside what the model can capture. The ionic limitation is shared with oxDNA's simplified screening.