Polymer Self-Similarity as a Resolution Probe for Replica-Averaged Ensembles

When scientists use simulation to reconstruct how a floppy, disordered protein moves, they run several copies of the molecule in parallel. They nudge the copies so that their average matches experimental measurements, which themselves average over billions of molecules. A second technique, metadynamics, pushes the copies over energy barriers so they explore more shapes. Every step of this pipeline can quietly distort the answer: too few copies, a poor choice of the coordinates being pushed, or a biased formula for predicting the measurement from a structure. Today there is no internal test that says which of these is distorting a given ensemble. Cosmologists faced the same problem with galaxy-formation simulations and solved it with a clean trick. In a universe with no built-in length scale, the true answer must look identical at every epoch once lengths are rescaled. Any failure of curves from different epochs to overlap therefore measures numerical error directly, with no external reference needed.

We propose to carry this trick into molecular ensemble modeling through a physical fact that has an exact analog there. Above a few residues, a disordered chain is a self-similar polymer. The typical distance between two residues grows as a power of how far apart they sit along the sequence, and the shape of the distance distribution becomes the same at every separation and every chain length once distances are rescaled. We will build a scale-free polymer testbed and use departures from this overlap to measure, and attribute, the errors that finite replica counts, chosen pushing coordinates, and correlated forward-model bias introduce into replica-averaged metadynamics. We will then derive transferable resolution rules for real disordered proteins.

Core Claim

For a chain in its scaling regime, let $s=|i-j|$ be the sequence separation between residues $i$ and $j$, $N$ the chain length, $R(s)=\langle r_{ij}^2\rangle^{1/2}$ the root-mean-square distance, and $x=r_{ij}/R(s)$ the rescaled distance. Self-similarity predicts

$$ R(s) \simeq b\, s^{\nu}, \qquad P_{s,N}(x) = f(x)\ \text{ for } s_0 \ll s \ll N, $$

Here $b$ is a monomer length, $\nu$ the chain's scaling exponent, $s_0$ the local-chemistry scale, and $f$ a single universal curve. We define the collapse residual $\Delta(s,N)=\int |P_{s,N}(x)-\bar f(x)|\,dx$, with $\bar f$ the pooled curve. This residual plays the role that $\Delta\xi/\xi$ across snapshots plays in cosmology.

The claim is that the main artifacts of replica-averaged inference leave distinguishable signatures in $\Delta(s,N)$. Finite replica number $M$ distorts the width and tails of $f$, and the distortion should shrink systematically as $M$ grows. Missing slow coordinates in the metadynamics bias break collapse at separations the bias does not drive. A correlated forward-model bias produces a smooth, separation-dependent offset in $R(s)$ that does not depend on $M$. Metainference's per-datum error parameters cannot see this last kind of error, because they treat errors as independent outliers.

This differs in kind from the cosmological work it borrows from. Quantifying resolution in cosmological N-body simulations using self-similarity and Good and proper test deterministic gravitational codes against time self-similarity. Here the self-similar variable is sequence separation, the system is stochastic, and the object under test is a Bayesian inference procedure, not an integrator.

Research Program

The work has three roles for the three directions. Self-similarity supplies the diagnostic. Replica-averaged metadynamics is the method under test. Residual systematic uncertainty, especially correlated forward-model bias, is the error class the diagnostic is built to expose.

The testbed. The testbed is a coarse-grained Cα chain with excluded volume, in the spirit of the Cα model used for denatured ACBP in Determination of an Ensemble of Structures Representing the Denatured State of the Bovine Acyl-Coenzyme A Binding Protein. We simulate it at lengths $N = 50, 100, 200$. Because the model is cheap, long unbiased sampling gives an exact reference ensemble. Collapse can be checked against this reference, which is the analog of a scale-free cosmology with known continuum behavior. From the reference ensemble we generate synthetic ensemble-averaged data: paramagnetic relaxation enhancement (PRE) from five spin-label sites, a hydrodynamic radius, and a sparse set of pairwise distances. PRE is a distance-sensitive NMR signal that averages $r^{-6}$, so its average is dominated by rare close contacts. The five-site layout mirrors the ACBP design.

Errors injected. We introduce three kinds of error, separately and in combination. The first is independent outliers, the regime metainference was designed for. The second is a correlated forward-model bias, such as a smooth miscalibration of the PRE distance conversion. The third is a deliberately wrong solvent quality in the force field, so that the restraints must fight the model. This follows the intentionally faulty force field in Metadynamic metainference.

The method under test. We run metadynamic metainference as implemented in PLUMED. Metainference (Metainference: A Bayesian inference method for heterogeneous systems) infers noise and systematic-error parameters alongside the structures. We scan $M$ from 4 to 64 replicas and three choices of pushing coordinates (radius of gyration only; radius of gyration plus end-to-end distance; a set of segment-level distances).

Mapping signatures to errors. For each run we compute $\Delta(s,N)$ and fit the three signature templates. We then ask whether the attributed error matches the known true error. This step turns the diagnostic into rules such as "the minimum $M$ at which PRE-restrained ensembles collapse to within a given tolerance at separation $s$."

The $r^{-6}$ prediction. We expect an analog of the two-body collisionality floor that Quantifying resolution in cosmological N-body simulations using self-similarity found scaling as $a^{-1/2}$. With finite $M$, the $r^{-6}$ average can be satisfied by forcing one replica into a close contact. That should produce a short-separation artifact whose extent shrinks with $M$, and we will measure its scaling empirically.

Transfer to real proteins. We apply the calibrated signatures to two real systems. The first is denatured ACBP, using PRE at 1.6, 1.9 and 3.0 M guanidine hydrochloride with a measured hydrodynamic radius of 25–26 Å (ACBP denatured state). The second is Aβ40, using chemical shifts from BMRB 17795 and hydrodynamic radii of 16.9 Å at 5°C and 15.5 Å at 40°C (The inverted free energy landscape of an intrinsically disordered peptide).

Real chains are not perfectly self-similar. ACBP has persistent contacts between residues ~15–35 and ~60–86. So on real data the diagnostic relies on the part of $\Delta$ that changes with $M$, with the pushing coordinates, or with the forward model, rather than on absolute collapse. Physical deviations from scaling should be invariant to those choices. This is the same logic by which scale-free cosmological results are transferred to ΛCDM, and it inherits the same semi-empirical caveat.

Baselines

The naive reference point is the convergence check used today. One check splits a trajectory in half and compares the halves, as in the trajectory self-similarity normalization of Comparing Molecular Dynamics Force Fields in the Essential Subspace, which applies a root-mean-square inner product to the top principal components. The other inspects one- or two-dimensional free-energy projections along the pushing coordinates, as in the Aβ40 metadynamics study. We will compute both on exactly the same trajectories as our diagnostic. Beating them means our diagnostic correlates with true ensemble error on the testbed, where the half-split and projection checks report "converged" while true error remains above tolerance.

The strongest existing method is metadynamic metainference itself. Its own inferred error parameters are its built-in diagnostic of systematic error. The file reports that metainference reached "2%±2%" ensemble error against "14%±5%" for maximum entropy under 30% outliers (Metainference). It also reports that M&M stayed within "< 0.06 $k_BT$" of perfect-data ensembles with 20% corrupted restraints (Metadynamic metainference). We run it with identical data, replica counts, simulation length and hyperparameter tuning budget. On the uncorrelated-outlier case we expect to match these numbers; that serves as a sanity check on our setup. The test of our proposal is the correlated-bias case. There, success means our diagnostic flags and correctly attributes the bias, whereas the inferred error parameters show no anomaly and the reconstructed ensemble error is materially above the outlier-only case.

Maximum-entropy reweighting serves as a third, cheaper baseline under the same data.

The key ablation removes the self-similar variables. We compute the same distribution-difference statistic on raw distances, without rescaling by $R(s)$ and without pooling across $s$ and $N$. If attribution accuracy survives this ablation, the scaling structure is not doing the work, and the proposal reduces to ordinary parameter scans.

Why not just…?

Why not just increase $M$ until the observables stop changing? This is the brute-force refinement that cosmologists found can show "apparent convergence" while the result remains physically unconverged (Quantifying resolution in cosmological N-body simulations using self-similarity). For pure finite-$M$ error it is a real threat and may work about as well as our diagnostic. The $M$ scan is in fact built into our design. It cannot, however, detect errors that do not depend on $M$, such as correlated forward-model bias or a missing slow coordinate. The testbed will show whether ensembles that are stable in $M$ yet biased exist. If they do not, we will narrow the claim to attributing coordinate and forward-model errors.

Why not just cross-validate on held-out restraints? Held-out data detects overfitting to noise. A correlated forward-model bias, however, contaminates training and held-out data identically, so held-out agreement stays good while the ensemble is wrong. We will run jackknife cross-validation on the correlated-bias testbed to show this directly. If it does flag the bias, this alternative wins, and we will report that.

Why not just trust metainference's inferred error parameters? They assume independent, per-datum or outlier-distributed errors (Metainference). A smooth bias across many data points can be absorbed into a modest global uncertainty, or into a distorted ensemble that fits it. The baseline comparison above tests exactly this.

Proposed Experiments

Decisive check: does the diagnostic track true error? On the testbed, across the full grid of $M$, pushing coordinates and error type, we compute the rank correlation between collapse residuals and true ensemble error. True error is measured as the distance between restrained and reference distance distributions and as the hydrodynamic-radius error. The proposal succeeds if this correlation clearly exceeds that of the half-split and projection checks and of the inferred error parameters.

Attribution test. We blind which error was injected and score how often the signature fit identifies it.

Finite-$M$ floor. We measure the smallest separation that converges, as a function of $M$ under PRE restraints, and test whether it follows a simple power law. It is not yet known whether such a law holds.

Ratio-cancellation test. This borrows the logic of the reactor baseline ratios in Model-independent short-baseline oscillations from reactor spectral ratios. We restrain on ratios of PRE profiles between chain lengths (testbed) or denaturant concentrations (ACBP), in self-similar variables, so that a shared multiplicative forward-model bias cancels. We then check whether collapse and true error improve relative to restraining on absolute values.

Transfer test. On ACBP and Aβ40 we report the $M$- and coordinate-dependent component of $\Delta$ and the replica count our rules recommend. We then check whether recommended-$M$ ensembles reproduce the held-out hydrodynamic radii better than the typical 8–20 replica setups.