Sacrificial Melting Sites as a Weight-Enumerator Problem in Supercoiled DNA

When a closed loop of DNA is underwound, as in a plasmid or a small circle, the strain has to go somewhere. Often it goes into a few base pairs that come apart, forming a "bubble" of single strands. The bubble releases twist and also acts as a floppy hinge. The striking fact is that this release is shared across the whole molecule. Simulations of 339-base-pair circles showed that inserting one easily melted, AT-rich stretch (the FUSE element) made that stretch open. The open stretch absorbed the strain and suppressed bubbles hundreds of bases away (Long-range correlations in the mechanics of small DNA circles under topological stress revealed by multi-scale simulation). This proposal starts from an observation about that effect. The statistical weight of a melting pattern depends on the sequence through local energies, and on everything else mainly through the total number of open bases. Mathematically, that is a sequence-weighted weight enumerator: a polynomial that counts configurations by their Hamming weight (number of nonzero entries), which coding theorists use to analyze error-correcting codes.

We propose to use this correspondence for an inverse-design question nobody has posed. How small and how weak can a deliberately placed "sacrificial" melting site be while still guaranteeing that the rest of a supercoiled molecule stays closed below a chosen probability? We will derive that guarantee with generating-function tools borrowed from coding theory, check it against coarse-grained simulation, and test the downstream claim that matters for DNA data storage in plasmids. The test is whether payload regions shielded by a sink suffer less strand-specific chemical damage, because damage is expected to be higher where DNA sits single-stranded.

Core Claim

In the standard two-state model of stress-induced melting (the SIDD calculation used by Sutthibutpong et al.), each base pair $i$ is either closed or open, $s_i \in \{0,1\}$. Opening costs a sequence-dependent free energy $\varepsilon_i(x)$ for sequence $x$, plus a penalty $a$ per bubble. Open bases relieve a linking deficit $\alpha<0$ through a residual-twist energy that depends only on the number of open bases $w(s)=\sum_i s_i$. The partition function therefore collapses to

$$Z(x,\alpha) = \sum_{w=0}^{n} N_x(w)\, \exp\!\left[-\tfrac{K}{2}\left(\alpha + \tfrac{w}{A}\right)^2\right],$$

where $K$ is the torsional stiffness constant, $A \approx 10.5$ is base pairs per helical turn, and $N_x(w)$ sums $\exp(-\sum_i \varepsilon_i s_i - a\cdot\#\text{bubbles})$ over all patterns of weight $w$. $N_x$ is exactly a weighted weight enumerator, and it can be computed as a polynomial by a transfer matrix. The stress-sink effect then becomes a statement about this polynomial. Suppose the low-$w$ coefficients of $N_x$ are dominated by patterns that open only inside a sink set $S$, and that $|S|$ can absorb $-\alpha A$ open bases. Then the global twist term no longer drives any site outside $S$, and the outside opening probability falls back toward its unstressed value. The design claim is a computable bound on $\max_{i\notin S} P(s_i=1)$ as a function of the sink's length and composition. The bound is obtained by comparing the enumerator restricted to $S$ with the full enumerator, in the same spirit as union bounds built from code weight spectra.

The coding-theory contribution is methodological. Bibak and Milenkovic evaluated weight enumerators of deletion-correcting codes defined by a single linear congruence. They did it with root-of-unity (Ramanujan and exponential) sums, which isolate a global constraint (Explicit Formulas for the Weight Enumerators of Some Classes of Deletion Correcting Codes). We reuse the idea that a single global constraint can be pulled out by a Fourier or Gaussian transform. Here the constraint is a quadratic penalty on the weight rather than a congruence. The quadratic can be written as a Gaussian integral over a uniform "torque" field, so the enumerator factorizes into independent per-site terms at each value of that field.

Relation to Prior Work

The physical motivation comes from three oxDNA and multi-scale studies.

That bending term depends on where a bubble sits, so it breaks the pure weight structure. We treat it as a position-dependent field, analogous to a complete weight enumerator, which tracks more than total weight. Whether that extension stays tractable is a genuine open question in this proposal.

Two results constrain the work. First, Sutthibutpong et al. report that oxDNA and SIDD both missed the twist-dependent breathing of a GC-rich DESIGNED sequence that atomistic simulation detected at $\Delta Lk \approx -0.5$. Any two-state enumerator inherits that blind spot, and we will report it rather than hide it. Second, Bensimon, Dohmi and Mézard showed that random static kinks in a chain only lower its effective persistence length at low force (Stretching a heteropolymer). Sequence disorder away from the sink should therefore enter only through averaged parameters, which supports keeping the non-sink model simple.

On the coding side, models of the DNA storage channel treat errors as independent and position-agnostic, as do constructions for burst deletions. Examples include profile-vector codes with GC and run constraints (Codes for DNA Storage Channels), array-based burst-deletion codes (Codes for Correcting a Burst of Deletions or Insertions), and empirical channel characterization on linear oligo pools (A Characterization of the DNA Data Storage Channel). None of these considers topology. This is reasonable for short linear oligos, which is why we restrict the storage claim to circular, supercoiled carriers.

Research Program

Analytical part. We will write $N_x(w)$ as a transfer-matrix polynomial with three states per site: closed, open, and bubble boundary. We will then derive the shielding bound and give its scaling with sink length $L$ and mean sink opening energy $\bar\varepsilon_S$. We expect protection to switch on sharply once $L \gtrsim -\alpha A$, because below that point the sink cannot carry the full deficit and the twist penalty keeps a nonzero torque on every outside site. Finally, we will add a positional bending field for circles under roughly 400 bp and measure how much it shifts the predicted bubble positions.

Computational part. We will rerun oxDNA with sequence-dependent parameters on the file's benchmark constructs: the 339 bp RANDOM, FUSE and DESIGNED circles, and the 600 bp block-random duplexes. From these runs we extract per-site opening profiles to serve as ground truth. We will then design sink-shielded variants by searching over sink sequences under the bound.

Experimental part. This tests whether the mechanics matter for stored information. We will build plasmid carriers with a fixed GC-balanced payload, with and without a designed sink, and age them following the accelerated-aging logic of the Grass and Heckel pools. We then sequence them and map the position of C-to-T errors relative to predicted opening probability.

Baselines

Unshielded payload (naive reference). The weakest reference is a GC-balanced, run-limited random payload with no sink, i.e. the RANDOM circle and its plasmid analog. It tells us what a supercoiled molecule does when nobody manages its melting. Beating it means the designed sink lowers the maximum outside opening probability, in oxDNA, by at least an order of magnitude at matched linking deficit.

Exact SIDD forward model with greedy search. This is the strongest existing method. Benham's SIDD calculation, as used by Sutthibutpong et al., is used to evaluate candidate sinks, and a greedy or simulated-annealing search picks the best one. It is the strongest competitor because it already computes the same two-state physics exactly. Fairness requires that it get the same candidate pool, the same evaluation budget measured in forward calls, and the same oxDNA evaluation. Beating it does not mean finding better sinks, since the two may tie. It means our bound certifies the result without search, reaching within 10% of SIDD-greedy's achieved suppression at a small fraction of the forward calls, and extending to the bending-coupled regime where SIDD has no term.

Independent-site ablation. This ablation isolates the key ingredient. We drop the global quadratic term, which turns the model into a local melting-propensity score close to GC content. It should fail to predict the suppression of distant bubbles in the FUSE circle. If it predicts oxDNA profiles as well as the full model does, the weight-enumerator view adds nothing.

Benchmark numbers. For the minicircle benchmark, there is no reported result on Lacuna for site-resolved opening probabilities or sink suppression ratios, so we will generate these numbers ourselves. For pinning, the reported figure is tip-bubble pinning of 34% to 100% of simulation time (Plectoneme tip bubbles). For aging, the reported figures are strand loss rising from 1.6% to 8% in high-redundancy pools and from 34% to 53% in low-redundancy pools after simulated aging (A Characterization of the DNA Data Storage Channel). These come from linear oligos, so they serve as context, not as a matched baseline.

Why not just…?

Why not just enforce GC balance and short homopolymers? This is what current storage codes do, and for linear oligos it may well suffice. In supercoiled DNA, Sutthibutpong et al. found that a GC-rich DESIGNED circle breathed more than an AT-rich one at $\Delta Lk \approx -0.5$. In the FUSE circle, a sequence choice made at one location changed opening hundreds of bases away. Both effects contradict any purely local composition rule. The independent-site ablation tests this directly. If local rules match the full model on oxDNA profiles, we drop the claim.

Why not just run SIDD or oxDNA inside an optimizer? This is a real threat for the pure twist case. The SIDD-greedy baseline is designed to settle it. If it matches our designs at comparable cost, the remaining contribution is the certificate: a closed-form guarantee and a scaling law for sink size. We will say so plainly. The case for the enumerator is stronger with bending coupling in small circles, and for constructs too long for oxDNA to run at scale.

Why not just store data in linear oligos, where topology is irrelevant? For most current pipelines that is the right choice, and we make no claim about them. The storage experiment applies only to plasmid or circular carriers, which are a natural format for in-cell or long-term archival storage. The physics result, a provable shielding bound for supercoiled molecules, stands independently of storage.

Why not just learn an empirical error model from sequencing data? A learned k-mer context model would capture local effects well. However, no dataset in the file comes from supercoiled carriers. A local context model also cannot represent the whole-molecule constraint by construction. In the aging experiment we fit such a model as a check. If it explains the positional pattern of damage as well as predicted opening probability does, mechanics is not needed for the error model.

Decisive Checks

Prediction. On the benchmark circles, the full enumerator must reproduce oxDNA's suppression of distant bubbles in the FUSE construct, and the independent-site ablation must not. We score this by Spearman correlation of per-site opening profiles and by the ratio of outside opening in FUSE versus RANDOM. We expect failure on DESIGNED at $\Delta Lk \approx -0.5$, matching the known gap between SIDD and atomistic simulation.

Design. Sink-shielded designs must keep outside opening in oxDNA below the predicted bound across a sweep of linking deficits. They must also show the predicted sharp onset of protection near $L \approx -\alpha A$.

Storage relevance. In aged plasmids, the C-to-T error rate within the payload should fall in shielded constructs relative to unshielded ones, and it should track predicted opening probability. A null result here, with the simulation results still holding, would establish that sink design controls melting but not storage damage, which is a useful outcome in its own right.