Chemists increasingly let a machine-learned energy model stand in for quantum chemistry. They train a neural network on quantum calculations, run millions of simulation steps with it, and then read off whether two nearby binding sites cooperate. Cooperation here means that the sites together hold a molecule more tightly than the sum of what each holds alone. The answer is a small difference between large numbers. In zeolites, pairs of neighboring acid sites can stabilize a second water molecule by up to −44 kJ/mol through a hydrogen-bonded "water bridge", by −11 kJ/mol in another framework, and not at all in a third (Water Adsorption at Pairs of Proximate Brønsted Acid Sites in Zeolites). Machine learning has a name for the worry this raises: predictive multiplicity, the finding that many models with the same overall accuracy give conflicting predictions for individual cases (Predictive Multiplicity in Classification). This proposal asks whether equally accurate energy models can disagree about whether cooperation exists at all.
The core hypothesis is that multiplicity concentrates in non-additive quantities. Training loss averages errors over configurations, and cooperative, bridged configurations are rare within that average. A second difference of energies also does not cancel errors that differ from one configuration to the next. Two potentials could therefore agree to within noise on every single-site adsorption energy and still disagree on the sign of the cooperativity. We propose to measure this by searching directly for the most disagreeing members of the set of near-optimal models, rather than by sampling random seeds. We then use the disagreeing models to decide where scarce high-level quantum labels should go.
Core Claim
Fix a trained energy model $h^*$ and a tolerance $\epsilon$. The Rashomon set $\mathcal{R}(\epsilon)$ is the set of models in the same architecture whose validation loss is within $\epsilon$ of $h^*$'s. For a two-site motif $m$, define the cooperativity functional
$$C_h(m) = E_h(m{+}2w) - E_h(m{+}w_a) - E_h(m{+}w_b) + E_h(m),$$
where $E_h$ is the model's energy, $m{+}w_a$ is the relaxed motif with one water at site $a$, and $m{+}2w$ is the motif with both sites loaded. $C_h(m)$ is the double-difference analogue of the extra-stabilization energy in Windeck et al. Our object of study is the cooperativity range $[\min_{h\in\mathcal{R}(\epsilon)} C_h(m), \max_{h\in\mathcal{R}(\epsilon)} C_h(m)]$. It carries over the viable prediction range of Dropout-Based Rashomon Set Exploration for Efficient Predictive Multiplicity Estimation from single outputs to a physically meaningful combination of outputs. Its classification counterpart is cooperativity ambiguity: the fraction of motifs whose label "cooperative" (say $C<-10$ kJ/mol) flips somewhere within $\mathcal{R}(\epsilon)$.
The claim has three parts. First, measured relative to its own scale, the cooperativity range is much wider than the range of single-site energies. Second, seed ensembles systematically underestimate it. Third, labeling the motifs where the extreme members disagree fixes cooperativity predictions faster than labeling where a committee's variance is highest.
The theory already hints at the second part. The midpoint anchoring identity in Model Agreement via Anchoring bounds disagreement between models in expectation over the training distribution. It says nothing about a rare subpopulation such as bridged motifs, and bridged versus non-bridged is exactly the sharp transition where learning curves need not flatten. Predictive Multiplicity in Classification showed on COMPAS that a model only 1% less accurate could change 17% of predictions, and that ad hoc retraining misses much of this. That result motivates our use of adversarial search over sampling.
No paper in the research file audits multiplicity in machine-learned potentials or in any cooperative binding quantity. Windeck et al. treat their potential as a single deterministic surrogate and check selected minima with CCSD(T), MP2 and dispersion-corrected DFT. That check validates the structures it happened to find. It does not show that a different, equally accurate potential would have found the same bridges.
Research Program
Training the reference models. We rebuild the zeolite setting of Windeck et al., covering H-MFI, H-FAU and H-CHA with bridged and non-bridged acid-site pairs. We train a standard message-passing interatomic potential on DFT-D energies and forces, and hold out a reference set of motifs that are relaxed and re-scored with MP2 and a CCSD(T) subset. The published −44, −11 and 0 kJ/mol values serve as sanity anchors. The raw dataset is not listed as a resource, so we regenerate it following their protocol.
Adversarial Rashomon search. For each target motif $m$, we fine-tune from $h^*$ to minimize $\pm C_h(m)$, subject to a penalty that keeps validation energy-and-force loss within $\epsilon$ of $h^*$. We set $\epsilon$ at 1% and 2% relative, mirroring the tolerance used by Marx et al. Every candidate is re-verified on the full validation set before it counts as a Rashomon member.
This is a gradient-based, non-exact stand-in for the exact mixed-integer search that is only possible for linear models. The two extremes it returns give a lower bound on the true cooperativity range, so any wide range we report is conservative. Re-running the MD sampling of Windeck et al. under each extreme model then tells us whether the water-bridge structures themselves appear or vanish.
Repairing the model. We follow the logic of Reimagining Meaningful Model Multiplicity, which shows that meaningful disagreement between $h^*$ and a subgroup-specific model reveals where $h^*$ is suboptimal, and that multiaccuracy-style correction removes it. Here the "subgroup" is the set of motifs where the extreme models' predictions of $C$ differ most. We send those motifs for quantum labels, patch $h^*$, and re-audit. This is the proposal's key ingredient: the disagreeing Rashomon members act as auditors that choose the labels.
Transfer to protein binding. To test whether the phenomenon is specific to zeolites, we repeat the audit on double mutants in SKEMPI 2.0, where the non-additive quantity is epistasis: the double-mutant ΔΔG minus the sum of the two single-mutant values. The synergistic, entropy-driven CI2 double mutant L49I/I57V escaped structure-based predictors (Double Mutant of Chymotrypsin Inhibitor 2 Stabilized through Increased Conformational Entropy). That case warns that non-additive targets are exactly where a single fitted model is least trustworthy.
Baselines
The naive reference point is a single potential trained once, with no uncertainty estimate. This is the workflow of Windeck et al. and of the affinity regressor in Unravelling the human taste receptor interactome. That regressor reached a Spearman correlation of only 0.19 on binding free energy, which is context for how weak single-model affinity prediction can be. It is not a number on our task. Against this baseline, the measure of success is how often its cooperative/non-cooperative call on a held-out motif disagrees with the quantum reference.
The strongest practical competitor is a deep ensemble, meaning independently seeded retraining, as used in committee-style uncertainty estimation. We run it with $M=10$ and $M=100$ models. The choice follows Measuring consistency via ensemble margin and local prediction variability, which reports that 10 models plus input perturbation reproduce a 100-model "warehouse", lowering the audit risk ratio on MRPC from about 0.4 to about 0.15. No reported result was found on Lacuna for seed ensembles, dropout, or committee disagreement on any cooperative binding energy, so the zeolite comparison has no published number to beat.
The second competitor is inference-time dropout exploration from Hsu et al., which reports 20x–5000x speedups over retraining on tabular and vision tasks. We add dropout to the same architecture and filter samples to $\mathcal{R}(\epsilon)$.
For fairness, all methods share the training data, architecture and $\epsilon$-verification step. The compute budget, counted in GPU-hours including the adversarial fine-tuning, is matched to that of the 100-model ensemble. Adversarial search beats these baselines if its verified cooperativity ranges are wider on a meaningful fraction of motifs and the widening holds up under quantum checking. That means at least one extreme model's prediction is actually closer to the reference than the ensemble mean on motifs where the ensemble was confident.
For the repair step, the ablation that isolates the key ingredient keeps the labeling budget fixed (for example 50, 100 and 200 MP2 labels) and changes only the acquisition rule. The rules are random, committee variance, dropout variance, variance of $C$ itself, and adversarial-range. Success means lower cooperativity error and lower post-repair ambiguity at equal budget.
Why not just…?
Why not just use a deep ensemble's variance, as practitioners already do? This is the most serious threat, and the experiment is built to settle it. Seed ensembles sample the typical output of training, not its extremes. Marx et al. found that ad hoc sampling underestimated worst-case discrepancy on COMPAS, and dropout explores only the basin around one optimum, as Hsu et al. acknowledge. Still, if the 100-model ensemble recovers most of the adversarial range and committee-variance labeling matches adversarial labeling, the adversarial machinery is unnecessary. We would then report the finding as "ensembles suffice for cooperativity auditing". That would still be a useful negative result, since nobody currently checks cooperativity this way.
Why not just add more quantum data everywhere? Scale alone has not fixed non-additive affinity prediction. The taste-receptor pipeline trained on millions of interaction records and still reached ρ = 0.19 on free energy. Globus-Harris and Garg argue that multiplicity tied to a structured subgroup persists until the subgroup is targeted. The random-acquisition arm of the ablation measures directly how much uniform extra data buys at matched cost.
Why not just skip the learned model and use CCSD(T) or long MD? Windeck et al. could find water bridges only because the learned potential made broad sampling affordable. High-level quantum chemistry was feasible only for re-scoring. The question is therefore not whether to use the learned model but how far to trust what it finds.
Why not just fit the cooperativity term directly with a correction model? A model trained to predict the correction term would face the same underspecification one level up, so we include it as an additional audited model rather than treat it as an escape from the problem.
Proposed Experiments and Decisive Checks
The first decisive check tests the central hypothesis. We compare the normalized cooperativity range, the range of $C$ divided by the reference-set spread of $C$, with the normalized range of single-site adsorption energies across all motifs in the three frameworks. If the two are comparable, the hypothesis that multiplicity concentrates in non-additive quantities is false.
The second check asks whether equally accurate potentials can overturn the qualitative conclusions. Would some verified member of $\mathcal{R}(\epsilon)$ predict cooperative bridges in H-CHA, or none in H-MFI? We re-run MD under the extreme models and re-score with MP2 or CCSD(T) whatever new minima appear.
The third check is the matched-budget acquisition comparison. Its metrics are mean absolute error of $C$ against the quantum reference, cooperativity ambiguity at 1% and 2% tolerance, and the flip rate of the cooperative/non-cooperative label.
The fourth check repeats the audit on SKEMPI 2.0 double-mutant epistasis with a standard graph-based ΔΔG regressor. It asks whether ambiguity is higher for epistatic pairs than for their single mutants.
A positive outcome is wide, verified cooperativity ranges on the first check together with an acquisition advantage on the third. That would mean conclusions about cooperative binding drawn from a single learned potential need a Rashomon audit before anyone synthesizes a material or molecule on the strength of them.
References
- Water Adsorption at Pairs of Proximate Brønsted Acid Sites in Zeolites
- Predictive Multiplicity in Classification
- Dropout-Based Rashomon Set Exploration for Efficient Predictive Multiplicity Estimation
- Model Agreement via Anchoring
- Reimagining Meaningful Model Multiplicity
- Double Mutant of Chymotrypsin Inhibitor 2 Stabilized through Increased Conformational Entropy
- Unravelling the human taste receptor interactome: machine learning and molecular modelling insights into protein-protein interactions
- Measuring consistency via ensemble margin and local prediction variability: Auditing decision systems in the presence of predictive multiplicity