When biologists measure how a toxic metal such as lead or cadmium shuts down an enzyme, they titrate the metal concentration and watch activity fall. The hard cases are not the dramatic ones — lead at 251.9 mg/kg crushing saccharification yield to 30% is unambiguous — but the borderline ones: cadmium trimming laccase-driven phenol removal by 27.6% in one soil region, arsenic seeming to do nothing up to 300 mg/L, or a neural network claiming near-perfect prediction of lead extraction from a handful of pot trials. These marginal claims are decided with the same tools as any garden-variety lab comparison: a few replicates, a t-test, a Bonferroni correction. Particle physics learned the hard way that this regime produces both false discoveries and false dismissals. When a parameter sits near a physical boundary, standard asymptotic tests break: the claimed sterile-neutrino signal in the Neutrino-4 reactor experiment fell from $3.2\sigma$ to $2.7\sigma$ once detector resolution entered the fit, and to roughly $2.2\sigma$ under exact Monte Carlo calibration, dissolving into background fluctuation (Neutrino-4 anomaly: Oscillations or fluctuations?).
Enzyme inhibition has the same structural problem. Relative catalytic activity is bounded between 0 and 100%, inhibition potency parameters are non-negative, and thresholds sit exactly where effect sizes are small and replicate counts are tiny — the conditions under which asymptotic statistics misbehave. Yet no paper in the retrieved literature applies calibrated, boundary-aware inference, blind analysis, or formal upper limits to metal-inhibition assays; the two literatures have never touched. This proposal imports the physics toolkit — parametric Monte Carlo significance calibration, profile likelihoods with boundary correction, masked analysis regions, and exclusion limits — into heavy-metal enzyme inhibition, and uses it to build a benchmark that tells the field which published marginal claims survive and which do not.
Core Claim
The claim is that nominal significance levels in heavy-metal enzyme inhibition assays are systematically overstated near inhibition thresholds, because activity is bounded and dose–response curves are fitted with asymptotic tests whose assumptions fail there. Replacing them with Monte Carlo–calibrated test statistics — the exact recipe that demoted the Neutrino-4 anomaly — will (a) reduce false claims of inhibition in true-null regimes, (b) recover genuine sub-threshold effects that Bonferroni-style tests miss, and (c) turn threshold estimates (minimum inhibitory concentrations) into statements with honest confidence intervals or upper limits instead of single sparse titration points. Formally, for a fitted inhibition model with parameters $\theta$ (including a bounded relative-activity scale $a \in [0,1]$ and a non-negative potency), we compute the profile-likelihood statistic
$$q_{\text{obs}} = -2\ln \frac{L(\hat{\theta}_{0})}{L(\hat{\theta})}, \qquad Z = \Phi^{-1}\!\left(1 - p_{\text{MC}}(q \ge q_{\text{obs}})\right),$$
where $\hat{\theta}_{0}$ is the best fit under the no-inhibition null, $L$ is the likelihood of the dose–response data, and $p_{\text{MC}}$ is obtained by generating many pseudo-datasets under the null and refitting — never by assuming $q$ follows a $\chi^2$ distribution, which is precisely what fails at the boundary.
Research Program
Three components. First, a synthetic benchmark generator for metal-inhibition assays: logistic dose–response curves in log metal concentration, with metal-specific potency, heteroscedastic assay noise, batch effects, and activity hard-clipped to $[0,1]$. Its parameters are calibrated so that synthetic experiments reproduce the documented effect sizes — lead dropping saccharification yield to 30% at 251.9 mg/kg while nickel leaves 87.4% yield at 10 mg/kg and the fungal cocktail retains over 95% filter paper activity in 100 mM Ni²⁺ (Metal accumulation by sunflower...); lead arresting hydrocarbon degradation above 100 mg/L while arsenic up to 300 mg/L leaves degradation efficiency constant (Microbial Removal of Petroleum Hydrocarbons...); and the 27.6% cadmium-specific laccase inhibition in the cathode region alongside 56.7% Cd²⁺ and 61.2% Cu²⁺ removal (Repairing Compound-Contaminated Soil Using EK-Laccase...). Because ground truth is known, we can measure the actual false-positive and false-negative rates of every competing test — something no real assay dataset can provide.
Second, a reanalysis of the documented marginal claims under the calibrated protocol. The arsenic arm of the A. baumannii JYZ-03 study is the key true-null case: efficiency held constant up to 300 mg/L As(V), so a correctly calibrated procedure must return no inhibition claim and instead a defensible upper limit on arsenic potency. Conversely, the cadmium-versus-copper cathode comparison in the EK-laccase work — a 27.6% difference in one soil region, with no reported replicate-level statistics on Lacuna for that comparison — is exactly the kind of bounded, single-region contrast where we expect naive and calibrated answers to diverge. Third, an overfitting audit of the machine-learning claim in Optimizing Sustainable Phytoextraction of Lead...: we generate pseudo-datasets from the authors' own quadratic response-surface model plus assay noise (a world with no hidden nonlinearity), train the same feed-forward back-propagation network with Levenberg–Marquardt optimization on the same sample sizes, and ask how often $R^2 \ge 0.9998$ arises from noise alone. This is the Giunti maneuver transplanted: an apparent high-significance fit checked against exact pseudo-data rather than asymptotic assumptions.
The protocol itself adopts blind analysis from the BABAR search for $B^0 \to \gamma\gamma$, where the signal region was masked until event selection and background models were frozen, eliminating experimenter bias before the unblinding (Search for the decay B0 -> gamma gamma). In our setting the masked region is the concentration band around a suspected threshold: the analyst fixes the dose–response model family, noise model, and test statistic on data outside the band, then unmasks. When no calibrated excess emerges, the output is an upper limit on inhibition potency rather than a hedged narrative claim. No reported result found on Lacuna for the numerical upper limit the BABAR analysis set (the file records only the Standard Model expectation of order $3 \times 10^{-8}$), so our protocol's limit-setting behavior is validated purely on the synthetic benchmark.
Baselines
The naive reference point is current practice: per-concentration replicate t-tests (or ANOVA) with a Bonferroni correction across the titration points, run at $\alpha = 0.05$. It is the strongest competitor in the sense that it is what essentially every paper in the retrieved bioremediation literature actually does, and beating it is the adoption hurdle. We run it on identical synthetic datasets and reconstructed titration grids, with the same replicate counts (three to five per concentration) the empirical studies use, and count as beating it a strictly lower false-negative rate at matched false-positive rate — the Bonferroni correction is expected to be especially blind to gradual, near-threshold inhibition.
The strongest existing method is the asymptotic profile-likelihood test using Wilks' theorem, the standard $\chi^2$ approximation that dominated before exact calibration. Neutrino-4 anomaly: Oscillations or fluctuations? reports this test's characteristic failure on bounded parameters: with the physical boundary and resolution effects included, Wilks gave 2.7σ where exact Monte Carlo gave about 2.2σ, against a naive 3.2σ. We run the Wilks test on the same fits and pseudo-datasets as our calibrated statistic, with identical models, data, and compute (the Monte Carlo step is the only addition, and it is cheap). Beating it means achieving the nominal false-positive rate — say, ≤ 5% at a nominal 95% threshold — where Wilks demonstrably exceeds it near the activity boundary, while losing minimal power in mid-range, unbounded regimes where Wilks is fine. For the overfitting audit specifically, the strongest competitor is the published neural network itself: Optimizing Sustainable Phytoextraction of Lead... reports a feed-forward back-propagation network at $R^2 = 0.9998$, predicting 86.05% lead extraction against an experimental maximum of 86.0%, versus a Box-Behnken response-surface model at $R^2 = 0.9033$. We retrain that architecture and training algorithm under our null pseudo-data; the network is beaten if its apparent accuracy collapses when the data-generating truth contains nothing beyond the quadratic model.
One ablation isolates the proposal's key ingredient: Monte Carlo calibration run with and without the bounded parameter space — that is, calibration on an unconstrained activity scale versus the true $[0,1]$ clipping. If boundary violation is the real driver of inflated significance, only the bounded version should restore nominal false-positive rates near thresholds. A second ablation compares blinded versus unblinded analysis on the same data, measuring how often analyst degrees of freedom (model choice made after seeing the threshold region) inflate the recovered significance.
Why not just…?
Why not just use t-tests and ANOVA with Bonferroni, as the field already does? The file contains direct evidence that this fails in both directions. In Mutual information analysis for feature and sensor subset selection..., differences between top sensor configurations were statistically insignificant at $\alpha = 0.01$ yet a real 80.6%-versus-79.6% performance ordering persisted downstream — small true effects that the standard test dismisses. And the Neutrino-4 reanalysis shows the converse failure: bounded parameters make asymptotic tests manufacture significance, with an apparent $3.2\sigma$ effect collapsing to roughly $2.2\sigma$ under exact simulation (Neutrino-4 anomaly: Oscillations or fluctuations?). Enzyme activity clipped to $[0,1]$ inherits both pathologies. Our synthetic benchmark settles this empirically by reporting false-positive and false-negative rates for t-tests against ground truth.
Why not just fit a response-surface model and read thresholds off the polynomial? The one head-to-head comparison in the file shows quadratic response-surface methodology structurally underfits these systems: the Box-Behnken model reached $R^2 = 0.9033$ with reported significance near 36%, while the true optimum involved high-curvature interactions among bacterial concentration, citric acid, and lead stress that the polynomial could not track (Optimizing Sustainable Phytoextraction of Lead...). More fundamentally, a fitted surface gives a point estimate of a threshold with no calibrated uncertainty, which is precisely what this proposal exists to fix. The response-surface model stays in our study as a data-generating model for the overfitting audit, so its role is honestly retained.
Why not just engineer hyper-tolerant strains or consortia so effect sizes are huge and statistics irrelevant? This is a real threat to motivation, and the evidence says it does not work: the resilient fungal consortium retaining over 95% activity in 100 mM Ni²⁺ still fell to 30% saccharification yield under 251.9 mg/kg Pb²⁺ (Metal accumulation by sunflower...); A. baumannii JYZ-03 shrugged off 300 mg/L As(V) but broke down above 100 mg/L Pb²⁺ (Microbial Removal of Petroleum Hydrocarbons...); and even after electrokinetic pretreatment removed over 56% of the cadmium, Cd²⁺ still suppressed cathode-region phenol removal by 27.6% (Repairing Compound-Contaminated Soil Using EK-Laccase...). Every tolerant system still has a transition zone, and deployment decisions (is this site's cadmium level safe for laccase treatment?) are made exactly inside that zone. Engineering away the statistics is not possible; making the statistics trustworthy is.
Why not just extend an existing enzyme benchmark with inhibitor-conditioned labels and train a bigger model? The natural ML reflex is to add metal co-solute features to a retrieval benchmark like ReactZyme, which reaches 35.88% Top-1 accuracy and 0.495 MRR on time-based splits over 178,463 enzyme–reaction pairs (ReactZyme: A Benchmark for Enzyme-Reaction Prediction), or CARE, whose contrastive pretraining reaches 96.7% Level-1 accuracy on easy splits but degrades badly on hard ones (CARE: a Benchmark Suite...), or the ECLIPSE hierarchical classifiers at 93.2% hierarchical F1 (Predicting Enzyme-Compound Interactions...). But no inhibitor-conditioned dataset exists on Lacuna — the P450 cascade work nearest to inhibition prediction handles organic compounds with binary thresholds, not metal ions with dose-dependent, bounded effects (Artificial neural network cascade identifies multi-P450 inhibitors...) — and any such dataset's labels would be threshold claims produced by the very uncalibrated assays this proposal audits. The calibrated protocol is the prerequisite, not the competitor: it is what makes trustworthy labels possible at scale.
Proposed Experiments and Decisive Checks
Experiment one runs the synthetic benchmark at documented effect sizes. Decisive metrics: false-positive rate at nominal $\alpha = 0.05$ for t-tests-with-Bonferroni, Wilks, and Monte Carlo calibration (prediction: the first two exceed nominal near thresholds, the third does not); false-negative rate at matched calibration (prediction: calibration beats Bonferroni on gradual inhibition); and coverage of minimum-inhibitory-concentration confidence intervals (prediction: naive intervals under-cover near the boundary). Experiment two applies the calibrated protocol to the reconstructed marginal cases: on the arsenic true-null arm it must return no claim plus an upper limit on As(V) potency; on the cadmium cathode comparison it must either confirm the 27.6% effect with calibrated significance or demote it, and either outcome is informative because the file records no replicate-level statistics for that comparison. Experiment three is the overfitting audit: if the Levenberg–Marquardt network reaches $R^2 \approx 0.9998$ on a meaningful fraction of null pseudo-datasets at the study's sample size, the published ANN-versus-RSM superiority claim is at least partly overfit assay variance; if it almost never does, the claim stands and the audit says so. The blinded-versus-unblinded and bounded-versus-unbounded ablations close the loop: if the unbounded ablation performs as well as the bounded one, the boundary-violation mechanism — and this proposal's central claim — is wrong, and we will report that.
References
- Neutrino-4 anomaly: Oscillations or fluctuations?
- Metal accumulation by sunflower (Helianthus annuus L.) and the efficacy of its biomass in enzymatic saccharification
- Microbial Removal of Petroleum Hydrocarbons from Contaminated Soil under Arsenic Stress
- Repairing Compound-Contaminated Soil UsingEK-Laccase Combination Repairing Technology
- Optimizing Sustainable Phytoextraction of Lead from Contaminated Soil Using Response Surface Methodology (RSM) and Artificial Neural Network (ANN)
- Search for the decay<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>B</mml:mi><mml:mn>0</mml:mn></mml:msup><mml:mo>→</mml:mo><mml:mi>γ</mml:mi><mml:mi>γ</mml:mi></mml:math>
- Mutual information analysis for feature and sensor subset selection in surface electromyography based speech recognition
- ReactZyme: A Benchmark for Enzyme-Reaction Prediction
- CARE: a Benchmark Suite for the Classification and Retrieval of Enzymes
- Predicting Enzyme-Compound Interactions for Enzyme-Catalysed Reactions
- Artificial neural network cascade identifies multi-P450 inhibitors in natural compounds