Enhancers are short stretches of non-coding DNA that switch genes on in particular tissues. Across related species they keep driving the same expression pattern even as their individual transcription-factor binding sites (the short motifs that regulatory proteins grab onto) are gained, lost and shuffled. Yeast ribosomal promoters and mouse–human enhancers show this conserved output despite heavy sequence change (Zeevi et al., Mouse ENCODE). This note proposes a quantitative explanation that can be tested. Fit a physical model of how proteins bind along the enhancer, then ask how far the effect of one binding site spreads through it. The answer is a "correlation length", the distance over which flipping one site measurably changes what happens at another. A site can then only be replaced by a site within that distance, and changes further away evolve independently.
The idea comes from the theory of correlation decay in statistical physics and approximate counting. There, when the influence between distant variables shrinks exponentially with distance, a global quantity can be computed exactly enough from a bounded local neighbourhood, with a proven error bound (Bandyopadhyay et al., Chen, Peng & Liu). We use this machinery for two purposes. Scientifically, it gives a sequence-specific, measurable length for each enhancer, and that length should predict where compensating turnover happens. Algorithmically, it gives an error-bounded window, which lets an evolutionary alignment model treat binding sites as interacting rather than independent without the state space blowing up.
Core Claim
Take a thermodynamic enhancer model of the kind in GEMSTAT, which is a statistical-mechanics model scoring every way transcription factors can occupy the enhancer, including cooperative binding and short-range repression (He et al., GEMSTAT). Once fit to expression data, it defines a Gibbs distribution over binding configurations, meaning a probability distribution in which each configuration's probability is proportional to its statistical weight. For a sequence $x$, define the influence of site $i$ on site $j$ as
$$I_x(i \to j) = \max_{s, s'} \left| P_x(\sigma_j \mid \sigma_i = s) - P_x(\sigma_j \mid \sigma_i = s') \right|,$$
where $\sigma_i$ is the occupancy state of site $i$ and $s, s'$ range over the states that site can take. Because the model is one-dimensional with finite-range interactions, influence must decay. It does so roughly as $\exp(-|i-j|/\xi_x)$, with $\xi_x = 1/\ln(\lambda_1/|\lambda_2|)$ computed from the ratio of the top two eigenvalues of the local transfer matrices. This form is the standard transfer-matrix result for 1D chains; for heterogeneous sequences we use the product of the local matrices.
The claim is not that decay exists; that is guaranteed in one dimension. The claim is that $\xi_x$ varies between enhancers and within them, that it is set by the fitted cooperativity and repression parameters rather than by any single fixed radius, and that it predicts two things. First, compensatory gain–loss pairs of binding sites across species fall within $\xi$ of each other more often than chance allows. Second, an evolutionary model with interacting sites, evaluated in $\xi$-windows, predicts regulatory modules better than an independent-site model.
Relation to Prior Work
The file reports no work that brings correlation-decay bounds to enhancer evolution, and the pieces sit in separate places.
GEMSTAT supplies the physical model. It computes the full partition function by dynamic programming along the enhancer. Adding cooperativity raised the average correlation between predicted and measured expression on 44 Drosophila segmentation enhancers from 0.516 to 0.587. The same work found that keeping only conserved sites hurt the fit. That finding motivates us, because functional sites are often lineage-specific, but GEMSTAT has no model of how they turn over.
EMMA supplies the evolutionary machinery (He et al., EMMA). It jointly aligns orthologous regulatory sequences and predicts modules, using a population-genetic substitution model that lets sites switch between functional and neutral states. It reaches an area under the ROC curve (AUC) of 0.95 at divergence 0.8. However, it treats binding sites as independent segments, which is exactly the assumption we relax.
Chen, Peng & Liu supply the method for bounding a local computation. They use Dobrushin's comparison theorem to bound the error of a marginal computed on a subgraph, and they grow the subgraph greedily. They tested this only on Ising grids and a citation network; we reuse the bound on sequence models.
CPModule (constraint programming) represents the hard-window alternative. It uses a fixed proximity window $\theta$, and the file reports that it is sensitive to that choice. The results reviewed in the file on correlation decay and CSP tractability (Correlation Decay and Tractability of CSPs, Montanari & Shah) stand as a constraint on our design. Guarantees hold only in the fast-decay regime, so we must report enhancers where $\xi$ is long and the local approximation fails rather than hide them.
Research Program
The first step is to measure $\xi$. We refit a GEMSTAT-style model on the 44-enhancer, 8-factor Drosophila segmentation dataset. We fit three variants: an additive baseline, one with cooperativity, and one with cooperativity plus short-range repression. For every enhancer and every position we compute the exact influence profile $I_x(i\to j)$. Exact computation is cheap here, because it needs only a pass of forward–backward dynamic programming (the standard two-sweep algorithm for chain models) for each conditioning. From the profile we extract a local correlation length $\xi_x(i)$ and check it against the Dobrushin contraction coefficient of the local conditional distributions. The coefficient gives a certified upper bound, and the exact profile gives the truth. This step also tells us how tight the theory is on biological parameters.
The second step is to test turnover against $\xi$. We align D. melanogaster and D. pseudoobscura orthologs of RedFly enhancers with EMMA. We call site gain and loss events from motif energy thresholds, as EMMA does. We then ask whether, for each lost site, a gained site with the same regulatory role (activator or repressor for the same target) lies within $\xi$ of the lost site's position more often than a position-shuffled null predicts. The comparison is against fixed-radius nulls of equal average size.
The third step builds an interacting-site evolutionary model. We extend EMMA's pair hidden Markov model so that the functional state of a site depends on its neighbours through the thermodynamic model. A joint state over all interacting sites would be exponential in their number. Instead we score each site's selective constraint by the change in predicted local occupancy within a window of $c\,\xi_x$, where $c$ is chosen so that the Dobrushin bound guarantees error below $\epsilon$. This keeps the state space bounded and makes the approximation error explicit, which is where correlation decay contributes something that 1D dynamic programming cannot provide in the joint-alignment setting.
Baselines
The naive reference point is the independent-site model with no locality structure. This is EMMA as published, together with a conservation filter that keeps only aligned, conserved sites. For module prediction, the strongest existing number is EMMA2's AUC of 0.95 at divergence 0.8 on its simulated benchmark, against 0.87 for Stubb and 0.82 for Morph (EMMA). On real RedFly/FlyReg data EMMA reports only "similar improvements," and we found no reported number on Lacuna for EMMA on real data. We will therefore rerun EMMA ourselves on identical sequences and splits.
The second baseline is fixed-window locality. It uses the same interacting-site model, but the window is either a single tuned radius or the fitted repression radius. We tune it on the same held-out enhancers with the same search budget as $c$, CPModule-style. Beating it means $\xi$-adaptive windows win at matched average window size.
The third baseline is the exact global thermodynamic model. It serves as the reference for how much accuracy truncation loses, and it is what $\xi$-windows must match to within $\epsilon$. On the segmentation data, GEMSTAT's reported average correlation of 0.587 with cooperativity is the target to reproduce with the truncated model.
Two ablations isolate the key ingredient. In the first, $\xi$ is computed from a model with cooperativity and repression switched off, so it reflects only motif density. In the second, $\xi$ values are shuffled across enhancers. If either ablation performs as well as the real $\xi$, then fitted interaction physics is not what drives the result.
All methods share alignments, motif matrices, simulation seeds and hyperparameter budget. Because EMMA's simulator generates independent sites, which would favour EMMA, we evaluate on two simulators. One is EMMA's original, where we must not lose accuracy. The other is a new simulator in which sites interact through the thermodynamic model and selection acts on predicted expression. We will report both, and we acknowledge that the second is our own construction.
Why not just…?
Why not just use a fixed window, such as the short-range repression radius? This is the real threat. If fitted $\xi$ is nearly constant across enhancers and roughly equal to the repression radius, the idea collapses into a known heuristic. The second and third steps test exactly this at matched window size. We will report the spread of $\xi$ across the 44 enhancers first, and if the spread is negligible we will say so and stop claiming adaptivity.
Why not just do in-silico mutagenesis with the full model, since 1D dynamic programming is linear time? For a single sequence we agree, and we compute exact influences this way ourselves. Correlation decay adds nothing to single-sequence speed. Its value lies elsewhere: it provides a certified locality length usable as a biological predictor, and it bounds the state space when interacting sites must be aligned jointly across species. That joint problem is where EMMA's independence assumption currently avoids a combinatorial blow-up.
Why not just use sequence conservation to find constrained sites? GEMSTAT already showed that conservation filtering degrades expression fits, because functional sites are often lineage-specific. The second step directly compares $\xi$-based compensation against alignment-conservation scores. The same logic applies to aligning first and scanning afterwards, which EMMA beat by 12–13% in site conservation sensitivity at divergence 0.7.
Why not sample interacting-site evolution with MCMC? This is feasible in principle. However, its mixing time is uncontrolled in strongly cooperative regimes, where the configurations are fractured into disconnected clusters (Montanari & Shah), and it gives no deterministic error bound. We will run a Gibbs-sampling version of the interacting model at equal compute as a secondary comparison.
Decisive Checks
Three results would settle the proposal. First, $\xi$ must vary meaningfully across enhancers, and it must shrink or grow with the fitted cooperativity and repression parameters. Second, compensatory gain–loss pairs must be significantly enriched within $\xi$ relative to both the shuffled null and the best fixed radius. Third, the $\xi$-windowed interacting model must beat rerun EMMA on the interacting simulator and on real RedFly enhancers, while matching it (within noise of 0.95 AUC) on EMMA's own simulator. It must also reproduce full-model expression fits within the stated $\epsilon$. If only the first check passes, we still gain a new descriptive statistic of enhancer architecture. If the second fails, correlation length does not constrain turnover, and we will report that negative result directly.
References
- Molecular dissection of the genetic mechanisms that underlie expression conservation in orthologous yeast ribosomal promoters
- A comparative encyclopedia of DNA elements in the mouse genome
- Counting without sampling: new algorithms for enumeration problems using statistical physics
- Efficient Localized Inference for Large Graphical Models
- Thermodynamics-Based Models of Transcriptional Regulation by Enhancers: The Roles of Synergistic Activation, Cooperative Binding and Short-Range Repression
- Alignment and Prediction of cis-Regulatory Modules Based on a Probabilistic Model of Evolution
- Cis-regulatory module detection using constraint programming
- Correlation Decay and Tractability of CSPs
- Counting good truth assignments of random k-SAT formulae