Computational drug design increasingly runs in layers. A cheap classical model of the protein and water is first used to simulate the system for a long time. A more accurate but expensive quantum or machine-learned model is then used to compute a correction to the binding free energy, which is the number that says how tightly a drug sticks to its target. The correction is usually small, a few kJ/mol. It is estimated from a limited number of expensive evaluations, and it is reported with an error bar that comes from the spread of those few samples. The open question is when that small correction has actually been measured, and when it is indistinguishable from sampling noise. Rare-event physics answers the same kind of question routinely. An experiment like EXO-200 publishes a median sensitivity, meaning the smallest signal it could expect to detect at its exposure, before claiming anything. Multi-level molecular simulation has no equivalent.
We propose to build one. The key observation is that the free energy correction between a cheap and an expensive model is an exponential average. Such an average is dominated by a few rare configurations where the two models disagree in a particular direction. Its precision is therefore set by an effective count of those configurations, and that count behaves like a Poisson count, so the relative error falls as one over the square root of the count. We will turn this into a working tool. From a small pilot set of expensive calculations, the tool predicts how many more calculations are needed to resolve a correction of a given size at a chosen significance. If the budget runs out, it reports an honest upper limit instead of a point estimate. It also serves as a stopping rule for the active-learning loops that decide when to call the quantum code.
Core Claim
Here is a concrete motivating case from the file. On the MCL1-19G protein-ligand benchmark, the machine-learned QM/MM workflow of Machine Learning-Enhanced Calculation of Quantum-Classical Binding Free Energies reports −35.3 ± 1.8 kJ/mol. That workflow trains a neural potential on quantum calculations of the drug region embedded in a classical protein (QM/MM) and uses active learning to choose which structures receive quantum calculations. The hierarchical embedding of Hierarchical Quantum Embedding by Machine Learning for Large Molecular Assemblies places a coupled-cluster core inside a density functional theory region inside molecular mechanics. It reports −37.2 ± 1.0 kJ/mol against an experimental −37.3 ± 0.1 kJ/mol.
The 1.9 kJ/mol shift between the two methods is smaller than their combined quoted uncertainty, which is about 2.1 kJ/mol. The more accurate method may well be better, but the published numbers do not establish that. Those quoted uncertainties are themselves sample spreads, and sample spreads are known to be over-optimistic when exponential averages are under-sampled. Our claim is that the resolvability of a level-of-theory correction can be predicted before the expensive calculations are run, by estimating how much the two models' Boltzmann distributions overlap. A Boltzmann distribution here means the probability each model assigns to each configuration at a given temperature.
Overlap can be measured with the order-2 Rényi divergence, a standard measure of how different two probability distributions are. Write $D_2$ for the order-2 Rényi divergence between the expensive-model distribution and the cheap-model distribution that actually generated the samples. The second moment of the reweighting factors equals $e^{D_2}$, an exact identity for importance sampling. This gives the number of independent expensive evaluations $n$ needed to resolve a correction of size $\delta$ at $z$ standard deviations, at temperature with thermal energy $k_BT$:
$$n \;\gtrsim\; \frac{z^2\,(k_BT)^2\,\left(e^{D_2}-1\right)}{\delta^2}.$$
The formula comes from propagating the relative variance of the mean weight, $(e^{D_2}-1)/n$, through the logarithm. The quantity $n/(e^{D_2}-1)$ is the effective number of configurations that carry the estimate, and it plays the role of the Poisson count $N$.
If the energy gap between the two models is Gaussian with spread $\sigma$, then $D_2 = \beta^2\sigma^2$, where $\beta = 1/k_BT$. In that case the required budget grows exponentially in the squared spread, not linearly. This is why a correction whose average is "relatively small", as the hierarchical-embedding paper describes its embedding shift, can still be expensive to resolve when its spread is not small.
The method layer of the proposal is supplied by Sensitivity analysis for rare events based on Rényi divergence, which bounds rare-event quantities under model perturbation using Rényi divergences. That work concerns stochastic differential equations, not molecular Hamiltonians. Our contribution is to use its bounds to put error bars on the estimate of $D_2$ itself. This matters because $D_2$ is the hard part: it is driven by the same rare tail that makes the correction noisy.
Research Program
The work has three parts.
The first part is a tail-robust estimator of the overlap. A naive plug-in estimate of $D_2$ from pilot samples is biased low, because the configurations that dominate it are the ones a small pilot set misses. We will bracket $D_2$ in three ways. The cumulant estimate assumes a Gaussian energy gap. The tail-fit estimate models the lower tail of the gap distribution with a generalized Pareto shape, which reflects the rare-event view. The Rényi-bound estimate uses the inequalities from the Rényi-divergence paper to give a conservative upper value. The sensitivity statement uses the conservative end of the bracket. The same machinery applies to non-equilibrium switching, the method both MCL1-19G papers use, in which the system is driven from the cheap to the expensive model over a short simulation. There the role of the energy gap is played by the dissipated work, and $D_2$ becomes a divergence between the forward and reverse path distributions.
The second part reports corrections the way physicists report weak signals. Each correction will come with a median expected sensitivity under the hypothesis that the correction is zero, mirroring EXO-200's practice of quoting a median sensitivity alongside its observed limit. The decision then follows a simple rule. If the observed correction exceeds the sensitivity, it is reported as a resolved shift. If not, it is reported as an upper limit, with no point value claimed.
The third part is a stopping rule for active learning. The workflow in Machine Learning-Enhanced Calculation of Quantum-Classical Binding Free Energies decides when to call quantum mechanics using a cutoff on the spread of an ensemble of machine-learned models, and it stops when the result appears converged. We will replace both decisions. Acquisition will target the configurations that dominate the estimated $D_2$, which are the tail configurations, rather than the configurations with the highest model disagreement. Stopping will occur when the predicted $n$ is reached or when the correction is bounded below a user-set relevance threshold, for example 1 kJ/mol. This directly addresses the risk the file raises: ensemble-variance sampling spends quantum calls on improbable outliers that contribute little to the free energy.
Baselines
The naive reference is the current reporting practice. A fixed budget of expensive evaluations or switching trajectories is run, and the correction is reported with a bootstrap or sample-spread error bar, as in both MCL1-19G papers. It will run with the identical budget and the identical cheap-model trajectories. Beating it means one of two things: intervals that achieve their nominal coverage when it does not, or nominal coverage at fewer expensive calls.
The strongest existing method on MCL1-19G is the hierarchical embedding of Hierarchical Quantum Embedding by Machine Learning for Large Molecular Assemblies. It reports a binding free energy of −37.2 ± 1.0 kJ/mol against the experimental −37.3 ± 0.1 kJ/mol. The strongest active-learning workflow is Machine Learning-Enhanced Calculation of Quantum-Classical Binding Free Energies, which reports −35.3 ± 1.8 kJ/mol. We do not aim to beat these accuracies with new physics. We use the same levels of theory and claim two things. First, our intervals are calibrated against a converged reference. Second, the active-learning workflow reaches equal or better calibrated precision with fewer quantum calls under our acquisition and stopping rule, at a matched total compute budget and with the same machine-learned potential architecture.
For the cheaper testbed we use the free-energy estimator from Enhancing Protein-Ligand Binding Affinity Predictions Using Neural Network Potentials. That estimator is UWHAM, an unbinned method that combines samples from many intermediate simulations, run within the Alchemical Transfer Method. The paper reports a CDK2 mean absolute error of 0.7 ± 0.1 kcal/mol for its neural-potential hybrid (ANI-2x with a classical protein force field) versus 0.8 ± 0.2 kcal/mol for the classical GAFF2 force field. We will ask whether the per-ligand corrections behind that 0.1 kcal/mol average difference are individually resolvable. For the ruthenium system GRP78-NKP1339 there is no reported numerical binding free energy on Lacuna, so it serves only as a stress test of heavy-tailed corrections.
Two ablations isolate the key ingredient. The first replaces the tail-robust $D_2$ estimate with the Gaussian cumulant estimate alone. The second keeps our estimator but restores ensemble-variance acquisition. If either ablation matches the full method, that ingredient is not doing the work.
Why not just…?
Why not just bootstrap more carefully and run longer? Bootstrapping resamples the configurations already collected. A tail configuration that never appears in the data can never be resampled, so bootstrap intervals for exponential averages shrink before the estimate has converged. This is a real threat if the actual corrections turn out to have nearly Gaussian gaps. In that regime bootstrapping and our Gaussian ablation should agree, and the contribution would reduce to budget planning. The coverage experiment below decides the question: we count how often bootstrap intervals contain the converged reference compared with ours.
Why not just run the expensive method directly and skip the correction? The file's evidence rules this out at the scale that matters. Coupled-cluster methods often lack analytical forces, so the hierarchical-embedding paper could train only on energies. Neural potentials need 0.5–1 fs timesteps, compared with 2–4 fs for classical force fields, as reported in QuantumBind-RBFE. The layered design exists precisely because direct sampling is unaffordable, and that is why its error bars need scrutiny.
Why not just use the ensemble-disagreement threshold that already works? It does appear to work on MCL1-19G. However, ensemble disagreement measures uncertainty in the learned model, not overlap between the two physical distributions. It can flag improbable configurations that carry almost no weight in the average while missing low-disagreement configurations that carry most of it. The acquisition ablation tests this directly. If ensemble variance matches tail-targeted acquisition in quantum calls needed for calibrated precision, we will report that the heuristic is adequate and keep only the reporting layer.
Proposed Experiments
The first experiment is calibration where ground truth is affordable. On CDK2 and TYK2 ligands from the shared benchmark subsets, we compute classical-to-neural-potential corrections with very long reference runs. Neural-potential calls are cheap enough to converge these corrections. We then draw many independent small pilot sets and test two things: whether the predicted $n$ matches the empirical number of samples at which intervals reach nominal coverage, and whether the intervals actually achieve 68% and 95% coverage. We repeat this with a semi-empirical and then a density functional theory upper level to span narrow and broad energy-gap distributions.
The second experiment is a verdict on MCL1-19G. We apply the sensitivity analysis to the density-functional-to-coupled-cluster embedding correction and to the full machine-learned QM/MM correction, and we state whether each is resolved or only bounded. The decisive outcome is whether the 1.9 kJ/mol difference between the two published methods exceeds the median sensitivity attainable at their reported budgets.
The third experiment compares compute at matched calibration. We rerun the active-learning workflow under our acquisition and stopping rule and count quantum calls until a calibrated ±1 kJ/mol interval is reached. The proposal succeeds if calibrated coverage holds where bootstrap coverage fails, and if quantum calls drop at equal calibrated precision. It fails cleanly if bootstrap intervals are already calibrated across all three testbeds.
References
- Search for Neutrinoless Double-<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mi>β</mml:mi></mml:mrow></mml:math>Decay with the Complete EXO-200 Dataset
- Machine Learning-Enhanced Calculation of Quantum-Classical Binding Free Energies
- Hierarchical Quantum Embedding by Machine Learning for Large Molecular Assemblies
- Sensitivity analysis for rare events based on Rényi divergence
- Enhancing Protein-Ligand Binding Affinity Predictions Using Neural Network Potentials
- QuantumBind-RBFE: Accurate Relative Binding Free Energy Calculations Using Neural Network Potentials