Confidence-Gated Side-Chain Repacking on Small Quantum Solvers

AlphaFold2 has a specific, documented calibration failure. At sites where a single amino acid has been mutated, it places the backbone almost perfectly but often gets the side chain (the part of the residue that sticks out from the backbone) wrong, while still reporting high confidence. Its per-residue confidence score, pLDDT, shows essentially no correlation with whether the side chain is right. Separately, variational quantum optimizers run a short tunable quantum circuit and let a classical computer adjust its parameters. They are limited to problems of roughly a hundred binary variables, so applying them to whole proteins is out of reach.

This proposal connects the two problems. A recalibrated, external error estimate decides which handful of residues around a mutation must be re-solved. Those residues become a small discrete side-chain choice problem that fits on today's quantum hardware. The spread of the resulting solutions is then turned back into a corrected confidence score. Put simply, calibration decides what goes on the quantum device, and the device's output repairs the calibration.

Core Claim

The claim has two parts, and only the first depends on nothing quantum. First, gating repacking by an external error estimate repairs side chains at variant sites and yields a confidence signal that tracks side-chain error. Gating by AlphaFold's own pLDDT will not work, and fixed-radius gating will be either wasteful or insufficient. Second, the gated subproblems are small enough that a variational quantum solver can be tested against their exact classical optimum. That makes the quantum step a falsifiable feasibility measurement rather than an asserted advantage.

At the subproblem sizes involved, exhaustive or near-exhaustive classical search is available. The best a quantum solver can do is match the classical optimum. The proposal is designed so that its biological conclusions stand regardless of how the quantum comparison comes out.

Relation to Prior Work

The motivating evidence comes from Analysis of AlphaFold2 for Modeling Structures of Wildtype and Variant Protein Sequences. On the S^sym variant set, that paper found sub-ångström backbones, side-chain RMSD up to 2.0 Å at mutation sites, and near-zero Pearson and Spearman correlation between pLDDT and side-chain accuracy. Those errors persisted after Amber relaxation. Two related papers show that confidence scores can be gamed or bypassed. Structural Memorization in AlphaFold shows overconfidence arising from template reliance. Inference-time optimization for experiment-grounded protein ensemble generation shows that ipTM can be inflated by a 0.01% embedding perturbation. Together these constrain the design: model confidence must never be the objective being optimized, only a gate or a prior, and the objective must be a physical energy.

Three papers supply methods. 3D-equivariant graph neural networks for protein model quality assessment introduced EnQA, an external network that scores structures independently. EnQA beats AlphaFold2's own pLDDT in per-model correlation with true lDDT, 0.7814 versus 0.7113. It was not validated on variant side chains, and it scores structures without fixing them. We use it as the gating signal and test whether it transfers to this setting.

Confidence-Weighted Elastic Gaussian Networks To Predict Protein Flexibility weights an elastic network's springs by pLDDT products and raises flexibility correlation from 0.765 to 0.841. Its principle, that confidence modulates physical coupling, inspires our confidence-weighted prior term. Its harmonic model, however, cannot represent discrete rotamer switches.

A Quantum Framework for Protein Binding-Site Structure Prediction on Utility-Level Quantum Processors ran VQE (the variational quantum eigensolver) on up to 102 qubits on IBM Eagle hardware. It folded short peptides on a coarse lattice with a two-stage scheme: first optimize the circuit, then sample it heavily and decode the best configuration. That paper discards side chains entirely and uses no confidence information, so it is the closest work and the clearest contrast. We reuse its two-stage execution on a different problem, side-chain choice on a fixed, trusted backbone.

Fragment molecular orbital-based variational quantum eigensolver for quantum chemistry shows that splitting a molecule into fragments keeps circuit width tied to fragment size. Its fragments come from chemical heuristics. Ours come from calibrated error estimates.

Research Program

The pipeline starts by predicting each wildtype and variant structure with AlphaFold2. An external estimator then predicts per-residue side-chain error. We use EnQA, optionally fine-tuned on a held-out split, with mutation identity and contact features as inputs. The residues whose predicted error exceeds a threshold, together with their contacting neighbors, form the free set $F$. Every other residue is frozen at its AlphaFold2 position.

Each residue $i$ in $F$ chooses one rotamer (a discrete side-chain conformation) from a library of $K$ candidates. This choice is encoded with binary variables $x_{ir}$, which equal 1 if residue $i$ takes rotamer $r$:

$$E(x)=\sum_{i\in F}\sum_r \big(e_{ir} - \lambda\, c_i\, \delta_{r,\hat r_i}\big)x_{ir} + \sum_{i<j}\sum_{r,s} e_{ir,js}\,x_{ir}x_{js} + \mu\sum_{i\in F}\Big(1-\sum_r x_{ir}\Big)^2 .$$

Here $e_{ir}$ is the energy of rotamer $r$ against the frozen environment, computed with an AMBER-type force field. The term $e_{ir,js}$ is the pairwise energy between rotamer $r$ at residue $i$ and rotamer $s$ at residue $j$. The prior term involves AlphaFold2's predicted rotamer $\hat r_i$; $\delta_{r,\hat r_i}$ equals 1 when $r$ is that rotamer. Its strength is set by $c_i$, the recalibrated confidence for residue $i$, and $\lambda$, a single global weight. The last term, with weight $\mu$, penalizes any assignment that does not pick exactly one rotamer per residue.

This energy is a QUBO (a quadratic objective over binary variables). The gating threshold is tuned so that $|F|\cdot K$ stays at or below about 100 qubits. For example, eight residues with twelve rotamers each give 96 qubits, matching the scale demonstrated on Eagle hardware. The QUBO is solved three ways: exactly, with branch-and-bound, which is feasible at this size; with simulated annealing; and with VQE, in simulation and on hardware.

The solver's sample distribution over low-energy assignments, taken from Boltzmann-weighted classical samples or from quantum measurement counts, gives a per-residue rotamer entropy. Combining this entropy with $c_i$ produces the output confidence.

Baselines

The naive reference is AlphaFold2 with Amber relaxation, reporting pLDDT as its confidence. Kabir et al. report side-chain RMSD up to 2.0 Å at mutation sites and near-zero pLDDT–accuracy correlation on S^sym. Any method must improve on both numbers.

The strongest existing scoring baseline is EnQA used alone, with no repacking. For global model ranking, the reported figure is per-model correlation 0.7814 on high-quality AlphaFold2 models. No reported result was found on Lacuna for EnQA on variant-site side-chain error. We will measure it on the same splits. Beating EnQA means achieving a higher Spearman correlation with side-chain error at mutation sites while also lowering RMSD, which EnQA cannot do by itself.

The strongest repacking baseline is ungated classical repacking. This means applying the same energy to every residue within a fixed radius of the mutation and solving it exactly or by annealing. No reported result was found on Lacuna for a classical rotamer packer on S^sym. We will run it with the same rotamer library, energy terms, and wall-clock budget.

For the quantum question, the comparison point is the exact classical optimum of the identical QUBO. VQE "beats" nothing here. It succeeds if it reaches the exact optimum on a stated fraction of instances, and if its sample entropies match the classical ones.

Three ablations isolate the key ingredients. The first gates by pLDDT instead of by the external estimate. The second sets $\lambda=0$, removing the confidence prior. The third adds the prior but gates with raw pLDDT. Together they separate the contribution of calibrated gating from that of confidence-weighted energy.

Why not just…?

Why not just solve the QUBO classically? At eight to ten residues, exact classical search is cheap and optimal, so this is a real threat to any claim of quantum benefit. We concede it in advance. The side-chain and calibration results are reported with the classical solver. The quantum arm measures whether noisy hardware, given confidence-sized subproblems, recovers the exact optimum and a faithful distribution over solutions. A negative answer is a publishable calibration of where today's devices stand, not a failure of the main claim.

Why not just run Amber relax or molecular dynamics? Kabir et al. already applied Amber relax to the top five AlphaFold2 predictions across all 684 mutations, and side-chain errors remained. Local minimization cannot jump between discrete rotamers. Full molecular dynamics per variant is described in the file's sources as taking days. We nonetheless include short restrained molecular dynamics on a 50-variant subset as a check.

Why not just filter with EnQA? Filtering identifies bad structures but does not produce good ones. EnQA's advantage over pLDDT is also documented only at the whole-model level. If EnQA turns out to be as blind as pLDDT at mutation sites, the gating step fails. Our first experiment tests exactly that, before any solver runs.

Why not just optimize AlphaFold3's latent space? Maddipatla et al. note that this kind of optimization is bounded by the model's prior. It also optimizes in the very space where confidence was shown to be gameable. Because its cost per variant is much higher than ours, we treat it as out of scope for the full set. We report it on a small subset only if compute allows.

Proposed Experiments

The data are S^sym, which contains 15 wildtype proteins, 342 variants, and 684 mutations including reverse mutations, split by wildtype family. Metrics are side-chain RMSD and chi-angle accuracy at mutation sites and their neighbors. We also report the Spearman correlation and expected calibration error between output confidence and side-chain error. As a secondary physical check, we compute the correlation of repacked energy differences with experimental ΔΔG (the change in folding stability caused by the mutation).

The first decisive check runs before any repacking. Does the external estimate correlate with mutation-site side-chain error where pLDDT does not? If it does not, the gating idea fails, and we report that. The second check asks whether gated repacking lowers side-chain RMSD relative to both Amber relax and fixed-radius repacking at equal compute. It also tests whether the entropy-adjusted confidence reaches a materially positive correlation with error, against a pLDDT baseline near zero. The third check is the quantum feasibility study. VQE runs in noiseless simulation, in noisy simulation, and on hardware at 32, 64, and about 96 qubits. We report the fraction of instances reaching the exact optimum and the total variation distance between quantum and classical sample distributions. Together these locate the size at which hardware noise, rather than problem difficulty, sets the limit.