DNA held in a closed loop cannot unwind freely. A minicircle, a plasmid, or a loop pinned by proteins has a fixed total "linking number," the count of how many times the two strands wind around each other. That budget can be spent on twist or on writhe, meaning the coiling of the helix axis in space. When a short stretch of the double helix transiently melts into a floppy single-stranded "hinge" (a breathing bubble), it spends some of that budget. Every other site in the loop then feels less stress. Simulations have already shown that one easily melted element can act as a "stress sink" that protects distant regions from melting, and that sharp kinks in stressed circles tend to form on opposite sides of the circle (Long-range correlations in the mechanics of small DNA circles under topological stress revealed by multi-scale simulation). What nobody has measured is the shape of this communication. Do two candidate hinges compete only for a shared twist budget, so that distance between them does not matter? Or do they also interact through the loop's three-dimensional bending geometry, so that the coupling depends on how far apart they are?
The question matters because the answer decides which models can be trusted. The first picture is exactly what the fast statistical-mechanics tool used for stress-induced melting (SIDD, which computes strand-separation probabilities over a whole closed domain) assumes. If that picture holds, SIDD is enough for genome-scale prediction. If the second holds, predicting where stress is released in real sequences requires geometry-aware models. This matters most for natural sequences, whose base composition is correlated over long stretches rather than random. We propose to measure a hinge–hinge coupling function across controlled families of correlated sequences. The measurement will show which picture holds, and in which regime of loop size and sequence correlation.
Core Claim
Let $h_i \in \{0,1\}$ indicate whether base pair $i$ sits inside an open hinge at a given instant, in a domain of $N$ base pairs held at fixed linking number. Define the connected coupling
$$C(d) = \big\langle \langle h_i h_{i+d} \rangle - \langle h_i \rangle \langle h_{i+d} \rangle \big\rangle_i ,$$
where the inner brackets average over thermal fluctuations, the outer average runs over positions $i$, and $d$ is the separation along the contour. The two pictures make different predictions. If hinges couple only through a shared twist budget, then beyond the width of a single bubble $C(d)$ is negative and flat in $d$: opening anywhere relieves everyone equally. If bending and writhe geometry also carry the signal, $C(d)$ has structure. In a sharply bent minicircle it should turn positive near $d = N/2$, the diametrically opposite point where the cooperative kinking of the multi-scale study appears. In a plectoneme it should show a specific signature: a tip bubble pins the tip locally but helps open sites half a loop away.
Our central claim is falsifiable. We expect geometric, distance-dependent coupling to dominate in domains shorter than about one persistence length (roughly 150 base pairs) and at plectoneme tips. We expect it to fade toward the flat, twist-only form in longer, weakly bent domains. We further expect the correlation length of the sequence to control which regime a natural sequence falls into. Sequences whose weak (AT-rich) stretches are longer than a bubble should behave as a single large sink. Sequences with short-range correlations should behave as many competing small sinks.
The contrast with prior work is specific. Earlier studies placed one engineered sink or one lesion in hand-built sequences. We measure the full pairwise coupling across statistically controlled sequence ensembles, and we test directly whether a twist-only model reproduces it.
Research Program
Coupled-hinge model. The main build extends the model that resolved the short-DNA cyclization puzzle. That work coupled the Czapla–Swigon–Olson (CSO) chain, a sequence-dependent 3D elastic model of the helix, to the Extended Peyrard–Bishop–Dauxois (EPBD) model of thermal base-pair breathing. It did so by switching a site to single-stranded elasticity whenever its opening crossed a threshold (Evaluating the role of coherent delocalized phonon-like modes in DNA cyclization). That work treated only open linear fragments with no topological constraint. We add a closed-loop linking-number constraint and a torque-and-force-clamped linear mode, so that opened hinges release twist and writhe consistently. Breathing is sampled with the GPU implementation of EPBD dynamics, which reports up to a 30-fold speedup over the C code (Efficient High-Throughput DNA Breathing Features Generation Using Jax-EPBD). Hinge rules are taken unchanged from the cyclization paper, so no parameter is tuned to the new targets.
Reference simulations. We use oxDNA2, a coarse-grained model with one rigid body per nucleotide, sequence-dependent stacking, and salt screening (Introducing improved structural properties and salt dependence into a coarse-grained model of DNA). It serves as the reference on a subset of sequences, reusing the 600 bp force-and-supercoiling setup of the plectoneme tip-bubble study (Plectoneme tip bubbles: Coupled denaturation and writhing in supercoiled DNA) and minicircles at the sizes of the multi-scale study.
Sequence ensembles. Here the heteropolymer physics enters as a design tool. Earlier theory showed that uncorrelated structural disorder, under weak stretching, merely softens the chain through a renormalized persistence length (Stretching a heteropolymer). Unzipping theory showed that disorder instead creates rugged barriers once strands separate (Unzipping of two random heteropolymers: Ground-state energy and finite-size effects). Neither addressed correlated disorder under topological stress. We generate three classes of sequences at matched GC content. The first uses Markov correlations with tunable correlation length $\ell$ from 2 to 200 bp. The second uses power-law correlated sequences. The third uses genomic windows, drawn from promoter regions like those where EPBD finds breathing hotspots (Toward a Detailed Description of the Thermally Induced Dynamics of the Core Promoter). The key control is a set of window-matched shuffles. These preserve the full distribution of 10-bp melting energies but destroy longer-range order. Any difference between a sequence and its shuffle is therefore an effect of correlation beyond the bubble scale.
Scope limit. The multi-scale study reports that at $\Delta\text{Lk} \approx -0.5$, atomistic dynamics shows breathing at CA, TA, CG, and AT steps in a GC-rich circle that neither oxDNA nor SIDD reproduces. Our model is built on EPBD and will likely inherit the same blind spot. We therefore restrict claims to stress levels where the coarse models agree with the atomistic results in that paper. We report the $\Delta\text{Lk} \approx -0.5$ regime separately as an unresolved test case rather than as a result.
Baselines
Naive reference. The simplest reference predicts hinges independently from each site's local melting free energy, with no coupling at all. This is equivalent to running linear EPBD per site. It predicts $C(d) = 0$ beyond one bubble width and sets the floor any coupled model must beat.
Strongest competitor: SIDD. The main competitor is the SIDD algorithm, as used in the multi-scale minicircle study. It is the natural, cheap, genome-scale embodiment of the twist-only hypothesis: every open base pair draws on one shared superhelical budget. It is also the tool practitioners would actually use. To keep the comparison fair, SIDD will receive the same sequences, the same $\Delta\text{Lk}$ or superhelical density, and the same salt condition. Its energy parameters stay at their published values, just as ours stay at the CSO-EPBD values, so neither side is tuned. The multi-scale study reports qualitative agreement across methods on the stress-sink effect but no scalar accuracy. There is no reported result found on Lacuna for SIDD against oxDNA on pairwise hinge coupling or plectoneme pinning. We will therefore establish SIDD's score ourselves.
Beating SIDD. Our model beats SIDD if two things hold on held-out sequences. First, it must match the oxDNA $C(d)$ curves with at least half the error. Second, it must predict the oxDNA pinning locus of the plectoneme significantly better.
Linear-regime anchor. The strongest numerical anchor comes from open fragments. CSO-EPBD predicted J-factors, the effective looping concentration measured in cyclization assays, within one order of magnitude of experiment for 82% of sequences under 100 bp. The rigid CSO model reached 25% on the 86-sequence benchmark (Evaluating the role of coherent delocalized phonon-like modes in DNA cyclization). Our extended model must recover about 82% when the constraint is switched off. Anything lower means the extension broke the physics.
Pinning target. For supercoiled linear DNA, the target is oxDNA's reported pinning fraction on block-random sequences, between 34% and 100% of simulation time depending on force (Plectoneme tip bubbles).
Ablation of geometry. The key ablation isolates the geometric ingredient. We run our own model with twist-release only, keeping hinge thresholds but freezing bending and writhe changes on opening. If this ablation matches the full model, geometry is not doing the work and the central claim fails.
Why not just…?
Why not just use SIDD, which already couples all sites through the linking constraint? This is the real threat. For long, weakly bent domains we expect SIDD to be adequate, and the proposal is designed to say so if it is. The evidence that it may fall short in tight or strongly bent loops is the cooperative kinking at opposite sides of minicircles in the multi-scale study. A distance-dependent effect of that kind cannot emerge from a single shared twist budget. The geometry ablation and the direct SIDD-versus-oxDNA $C(d)$ comparison settle this. If SIDD's flat coupling matches oxDNA across the tested loop sizes, we will report SIDD as sufficient and redirect the sequence-correlation study onto SIDD.
Why not just run oxDNA on every sequence? oxDNA is the reference, but the ensemble cannot be run with it. Each correlation class needs hundreds of sequences, several loop sizes, and several stress levels, all sampled well enough to estimate pairwise covariances. The published tip-bubble study itself examined only three sequence architectures at 600 bp. We use oxDNA on a stratified subset of roughly 5–10% of the sequences, which is enough to validate or reject the fast model.
Why not just predict the pinning site as the weakest local window? Extreme-value reasoning says the plectoneme sits at the lowest-melting window, regardless of what surrounds it. The window-matched shuffles test exactly this. If shuffled and original sequences give statistically identical pinning loci and $C(d)$, the local predictor is sufficient and correlation beyond the bubble scale is irrelevant. We would report that as a negative result.
Why not just renormalize the persistence length, as heteropolymer theory suggests? That result was derived for weak stretching without melting. It is already contradicted in the strongly bent regime: worm-like chain models underestimate short-fragment cyclization by up to six orders of magnitude, and body kinking replaces smooth bending below about 80 bp (Coarse-grained modelling of strong DNA bending II: Cyclization). We keep a renormalized-chain curve only as a large-domain limit check.
Proposed Experiments
Regression check. First, reproduce the 86-sequence J-factor benchmark with the constraint removed and confirm recovery of about 82%.
Minicircle coupling. Second, in minicircles of 90, 150, and 300 bp at several negative $\Delta\text{Lk}$ values, compute $C(d)$ with the full model, the geometry ablation, SIDD, and oxDNA. The decisive observation is whether oxDNA's $C(d)$ has a positive lobe near $d = N/2$ that the full model captures and SIDD misses. We also test whether that lobe shrinks with circle size, as the claim predicts.
Correlated sequences under supercoiling. Third, in the 600 bp force-and-supercoiling setup, sweep sequence correlation length $\ell$ against window-matched shuffles. We measure three quantities: the pinning fraction, the entropy of the pinning-locus distribution, and suppression of distal breathing. The prediction is a crossover in pinning behavior as $\ell$ passes the tip-bubble size. The shuffles should show no such crossover.
Lesion as a fixed sink. Fourth, a lesion test. Place a pre-bent defect of the thymine-dimer type in correlated backgrounds and measure how it alters $C(d)$ and distal breathing. That defect lowers the plectoneme threshold to $\sigma \approx -0.038$ in random backgrounds (The interplay of supercoiling and thymine dimers in DNA). The test asks whether a permanent hinge silences transient ones elsewhere in the loop. That would matter for proteins such as Fis, which read breathing rather than bases (Binding of Nucleoid-Associated Protein Fis to DNA Is Regulated by DNA Breathing Dynamics).
Outcomes. The study succeeds if two results hold together. SIDD's flat coupling must fail in tight loops and succeed in long ones. And the geometry ablation must fail wherever SIDD fails. The central claim is refuted if SIDD matches oxDNA everywhere, or if shuffles are indistinguishable from correlated sequences.
References
- Long-range correlations in the mechanics of small DNA circles under topological stress revealed by multi-scale simulation
- Evaluating the role of coherent delocalized phonon-like modes in DNA cyclization
- Efficient High-Throughput DNA Breathing Features Generation Using Jax-EPBD
- Introducing improved structural properties and salt dependence into a coarse-grained model of DNA
- Plectoneme tip bubbles: Coupled denaturation and writhing in supercoiled DNA
- Stretching a heteropolymer
- Unzipping of two random heteropolymers: Ground-state energy and finite-size effects
- Toward a Detailed Description of the Thermally Induced Dynamics of the Core Promoter
- Coarse-grained modelling of strong DNA bending II: Cyclization
- The interplay of supercoiling and thymine dimers in DNA
- Binding of Nucleoid-Associated Protein Fis to DNA Is Regulated by DNA Breathing Dynamics