When DNA is bent sharply, at the tip of a supercoiled loop or around a circle shorter than about 100 base pairs, it often stops behaving like a smooth elastic rod. A few base pairs come apart, and the short single-stranded patch acts as a floppy joint, a "transient hinge." DNA is also not uniform: its sequence varies along the chain, so some places are much easier to melt than others. This note proposes treating the hinge as a particle sitting on a bumpy one-dimensional energy track, where the bumps come from the sequence. The statistics of those bumps can be computed from the sequence correlation function. If the hinge is confined, as in a tiny circle, it settles in the deepest valley of the track. If it can travel, as at the tip of a plectoneme (the twisted, interwound loop that supercoiled DNA forms), it hops between valleys, and its effective diffusion constant falls exponentially with the roughness of the track. One quantity, the variance of the sequence-induced hinge energy over a window the size of the hinge, should then predict both where DNA kinks and how strongly kinks get stuck.
This matters because the two literatures have so far talked past each other. Heteropolymer physics has detailed theory for chains with frozen ("quenched") sequence disorder: rugged landscapes, glassy phases, and exact scaling of unzipping fronts. Hinge studies have strong simulations and benchmarks but test only a few hand-built sequences. Genomic DNA has correlated sequence structure at many scales. Whether supercoiling-driven loops migrate freely along real DNA, or get trapped at particular sites such as promoters, depends on exactly the link between these two bodies of work that has not yet been made.
Core Claim
Write $U(x)$ for the free energy of the best hinge configuration centred near position $x$: the cost of melting and unstacking a bubble, minimised over bubble sizes compatible with the imposed bend or supercoil. For a hinge that can move, the classic result for diffusion in a one-dimensional periodic or stationary random potential gives
$$ \frac{D_{\text{eff}}}{D_0} = \frac{1}{\langle e^{U/k_BT}\rangle\,\langle e^{-U/k_BT}\rangle} \;\approx\; e^{-\sigma_U^2/(k_BT)^2}, $$
where $D_0$ is the hinge's diffusion constant on a uniform sequence, the angle brackets average along the chain, and the approximation holds when $U$ is roughly Gaussian with variance $\sigma_U^2$.
For a hinge that is confined, as in a short circle of length $L$, the hinge sits at the minimum of $U$. For a landscape with correlation length $\xi$, extreme-value statistics put that minimum near $-\sigma_U\sqrt{2\ln(L/\xi)}$.
The sequence enters only through $\sigma_U$ and $\xi$. With $\epsilon$ the AT/GC energy contrast and $C(r)=\langle s_i s_{i+r}\rangle$ the composition correlation function of the sequence, a crude fixed-window version is $\sigma_U^2 \approx \epsilon^2\sum_{i,j\le w}C(i-j)$, where $w$ is the hinge size.
Two things in this claim are new. The first is a single law linking the correlation structure of a sequence ensemble to both hinge placement and hinge mobility. The second is the prediction that the hinge is soft: it can shift and resize inside the bend, so its landscape is smoother than a naive fixed-window melting profile. That second prediction is what separates the proposal from the obvious "find the AT-rich window" heuristic.
Grounding in Prior Work
Work that motivates the gap. Plectoneme tip bubbles used the oxDNA coarse-grained model to show that bubbles nucleate at plectoneme apices and pin the plectoneme for 34% to 100% of the trajectory, depending on force. It tested only uniform, random (49% GC) and 120-bp block-random sequences, and offered no theory linking sequence statistics to pinning. Long-range correlations in the mechanics of small DNA circles combined atomistic simulation, oxDNA and the stress-induced duplex destabilisation (SIDD) algorithm on three handcrafted minicircles. It found that an easily melted element acts as a "stress sink" that suppresses melting elsewhere. In our language, the hinge occupies the global minimum of $U$, and stress relief is the topological reason only one hinge forms. That paper also found that a GC-rich repetitive sequence breathed more than expected because of its intrinsic twist. This constrains us: $U$ cannot be built from base-pair composition alone and must include twist and stacking terms.
Work that supplies the method. The theory of quenched disorder is the methodological lens. Unzipping of two random heteropolymers showed that disorder turns a smooth front into jump-like, pinned motion with $n^{1/3}$ energy fluctuations. Glassy phases in random heteropolymers with correlated sequences showed that Markovian sequence correlations change the qualitative phase structure. We reuse the logic of both but not their geometry: a single hinge moving along a chain is a much simpler object than two strands unzipping or a collapsed globule. That simplicity is what makes exact transport formulas usable. We also reuse the EPBD breathing model (a nonlinear model of base-pair opening driven by thermal fluctuations) through Jax-EPBD and the EPBD surrogate. Both make genome-scale opening profiles cheap.
Work that constrains the design. Stretching a heteropolymer showed that, at low force, disorder simply renormalises the chain into a softer homopolymer. Our claim is that this smoothing fails once the chain must make a hinge, because a hinge is a localised object that samples the extremes of the disorder rather than its average. The role of structural parameters in DNA cyclization found that harmonic sequence-dependent elasticity matches only 24% of sequences shorter than 70 bp. Coarse-grained modelling of strong DNA bending II found that below about 80 bp oxDNA kinks are forced by bending stress. Together these support a model in which stress decides where the hinge opens. That is the mechanical-to-thermal feedback missing from the one-way CSO-EPBD coupling in Evaluating the role of coherent delocalized phonon-like modes in DNA cyclization.
Research Program
The program has three parts.
Building the landscape. For a sequence, an imposed geometry and a supercoiling level, we compute $U(x)$ in two ways:
- Fixed-window profile. Nearest-neighbour melting free energy averaged over $w$ base pairs.
- Adaptive hinge profile. The free energy minimised over bubble length (1 to about 12 bp) and offset within the bent region. It includes the bending energy released by the hinge and a sequence-dependent twist and stacking term, so that the GC-twist anomaly can appear.
We calibrate the adaptive profile against oxDNA free energies obtained by restraining the plectoneme apex at a handful of positions with umbrella sampling, as used in the cyclization study.
Building sequence ensembles with controlled correlations. All ensembles are held at 50% GC:
- Markov chains with tunable AT/GC persistence, giving correlation lengths from 1 to 50 bp.
- Power-law correlated sequences.
- Block-random sequences that reproduce the Matek design.
- Composition-preserving shuffles of each of the above.
- Natural human sequences drawn from chromosomes 8 and 9, the held-out split of the EPBD surrogate dataset.
The design deliberately pairs ensembles with equal predicted $\sigma_U$ but different $\xi$, and ensembles with equal GC fraction but very different $\sigma_U$.
Testing the two regimes. In the mobile regime we run oxDNA on supercoiled duplexes under tension, following the 600 bp protocol and extending to about 1.2 kb so that plectoneme hopping between valleys is observable. We measure the tip position over time, the pinning fraction and $D_{\text{eff}}$. In the confined regime we replace the CSO-EPBD hinge rule with a stress-selected hinge: a kinked chain whose single hinge sits at the minimum of the adaptive $U$. We score it on the 86-sequence cyclization benchmark.
We then apply the cheap landscape to the whole genome to map $\sigma_U$ and predicted pinning sites. Using the core-promoter work of Toward a Detailed Description of the Thermally Induced Dynamics of the Core Promoter, we test whether transcription start sites are enriched for deep hinge traps.
Baselines
Uniform reference. The naive point of comparison is the sequence-averaged chain: the "average-base" oxDNA sequence, with plectoneme diffusion $D_0$ and zero pinning, together with Bensimon-style renormalised homopolymer elasticity. Every disordered ensemble is compared against this at identical force, supercoiling, length and simulation time. Beating it simply means a measurable reduction in $D_{\text{eff}}$ or a non-zero pinning fraction. The oxDNA tip-bubble paper already establishes this qualitatively, with pinning of 34% to 100%. The quantitative result we must add is a collapse onto the predicted $\sigma_U$ dependence.
Fixed-window melting profile. This is the strongest cheap competitor for the plectoneme question. It predicts that the hinge sits in, and is pinned by, the most AT-rich window. It is computed from exactly the same sequences, with $w$ tuned on a held-out ensemble using the same budget given to the adaptive model's calibration. The proposal beats it if the adaptive landscape predicts $\log D_{\text{eff}}$ and pinning location with clearly lower error on held-out ensembles. The effect should be largest where the two disagree: short-correlation sequences, where a soft hinge averages away small bumps, and twist-anomalous repeats. If both perform equally, the adaptive hinge ingredient is unnecessary and we will report that.
CSO-EPBD for cyclization. The strongest existing method on the confined question is CSO-EPBD. It predicts J-factors (the effective concentration that measures how readily a chain's ends meet) within one order of magnitude of experiment for about 82% of sequences shorter than 100 bp, against 25% for unhinged CSO, as reported in Evaluating the role of coherent delocalized phonon-like modes in DNA cyclization. For sequences shorter than 70 bp, the best harmonic parameter sets reach 24% (The role of structural parameters in DNA cyclization). We score our stress-selected hinge model on the same 86 sequences, with the same within-one-order-of-magnitude criterion and the same structural parameter set. We fit no per-sequence parameters.
Matching 82% with a single-hinge extreme-value rule would show that stress-selected placement captures what thermal breathing captures. Exceeding it on the shortest fragments would support feedback over one-way coupling. Falling clearly below it would mean the global-minimum picture misses multi-hinge or breathing contributions. The benchmark mixes assay types, so we also report ligation-based and FRET-based subsets separately; the oxDNA cyclization study shows the two systematically differ.
Correlation ablation. The key ablation isolates sequence correlation itself. Each correlated ensemble is compared with its composition-preserving shuffle. This tests whether $C(r)$, beyond GC fraction, controls mobility as predicted through $\sigma_U$.
Why not just…?
Why not just pin the hinge to the most AT-rich window? This is a real threat. In the deep-trap limit our theory reduces to it. The difference is that the hinge adapts its size and position, which predicts systematically smaller effective roughness for short-correlated sequences, plus a twist correction that composition alone cannot provide. The minicircle study already shows that composition alone mis-ranks the GC-rich repeat. The fixed-window baseline above is designed so that this alternative can win, and the held-out ensembles will show whether it does.
Why not just run oxDNA on natural sequences? oxDNA is our ground truth, but it is expensive at kilobase scale, and a handful of trajectories gives no rule that transfers to new sequences. The proposal uses oxDNA to calibrate and test a landscape that can then be evaluated on a million genomic windows in minutes with the EPBD surrogate. The test is whether oxDNA-measured $D_{\text{eff}}$ collapses onto $\sigma_U$. Without such a collapse, more simulation would just be more anecdotes.
Why not just use a renormalised or harmonic sequence-dependent elastic model? The renormalisation in Stretching a heteropolymer applies to smooth deformations. Harmonic sequence models match only 24% of sub-70 bp circles, and oxDNA shows forced kinking below about 80 bp. A localised hinge responds to extremes of the disorder, not its average, so averaged elasticity cannot predict pinning. Our uniform-reference comparison makes this failure explicit.
Why not just add stress feedback to CSO-EPBD? That is a worthwhile and complementary project. Our cyclization arm is effectively its simplest version, with stress choosing the single weakest site. A feedback-coupled CSO-EPBD would still be a per-sequence calculator, however. It would not by itself say how correlation length controls mobility across ensembles, and that is the question the plectoneme arm answers.
Decisive Checks
The plectoneme arm succeeds if three things hold across Markov, power-law, block and natural ensembles:
- $\ln(D_{\text{eff}}/D_0)$ from oxDNA falls on one line against $\sigma_U^2$ from the adaptive landscape, with slope near $-1/(k_BT)^2$.
- Ensembles matched in $\sigma_U$ but differing in $\xi$ share the same long-time $D_{\text{eff}}$. They should differ instead in sample-to-sample scatter and in how long they take to converge. This is the signature of long-range correlations making the hinge's behaviour non-self-averaging, the analogue of the glassy regimes in the correlated-heteropolymer theory.
- Shuffled controls lose the predicted slowdown.
The cyclization arm succeeds if the stress-selected hinge reaches accuracy comparable to the reported 82% on the sub-100 bp subset without per-sequence fitting. A further prediction is that the model's residuals are concentrated in sequences whose landscape has two near-degenerate minima, where a single-hinge rule should fail.
The genome map is a forward prediction. It is meaningful only if the first two arms hold. It states which promoter classes should trap supercoiling-driven loops, a claim that magnetic-tweezers experiments could test directly.
References
- Plectoneme tip bubbles: Coupled denaturation and writhing in supercoiled DNA
- Long-range correlations in the mechanics of small DNA circles under topological stress revealed by multi-scale simulation
- Unzipping of two random heteropolymers: Ground-state energy and finite-size effects
- Glassy phases in random heteropolymers with correlated sequences
- Efficient High-Throughput DNA Breathing Features Generation Using Jax-EPBD
- Scalable DNA Feature Generation and Transcription Factor Binding Prediction via Deep Surrogate Models
- Stretching a heteropolymer
- The role of structural parameters in DNA cyclization
- Coarse-grained modelling of strong DNA bending II: Cyclization
- Evaluating the role of coherent delocalized phonon-like modes in DNA cyclization
- Toward a Detailed Description of the Thermally Induced Dynamics of the Core Promoter