Two-Channel Attribution of Antioxidant Activity to GC-MS Peaks

Food chemists routinely run two experiments on the same plant oil or fermented drink. The first is gas chromatography–mass spectrometry (GC-MS), an instrument that separates a vapor into its individual compounds and identifies each one by how it shatters. The second is a color-change test for antioxidant capacity, which measures how much of a colored "radical" chemical a sample neutralizes. They then try to say which compounds cause the activity, and they mostly fail. The two measurements are made on separate whole-sample portions. Different color tests disagree with one another. The final link between compound and activity is a correlation coefficient computed over a handful of samples.

We propose to treat this as a joint inference problem, borrowing two design ideas from neutrino physics. Neutrino physicists faced a structurally similar problem: they had to explain an apparent signal using several imperfect, mechanism-specific measurements whose biases do not cancel. Their answer was to combine channels that share underlying parameters and to compare a near and a far detector so that common biases drop out. We do the same here. Each GC-MS peak gets two latent capacities, one per reaction mechanism. All color tests are fitted at once under that shared structure. A matched "stripped" portion of the sample, with its volatiles removed, acts as the near detector that cancels matrix and optical artifacts.

Core Claim

Attributing antioxidant activity to individual volatile compounds is currently underdetermined for two reasons. A single assay mixes chemistry with probe-specific artifacts, and the non-volatile matrix contributes signal that GC-MS cannot see.

Our first ingredient is a mechanism-structured joint fit. Every compound $i$ gets a hydrogen-atom-transfer capacity $h_i$ (the antioxidant hands a hydrogen atom to the radical) and an electron-transfer capacity $s_i$ (it hands over an electron). Every assay $a$ responds to a sample $k$ as

$$y_{ak} = \sum_i c_{ik}\,(\alpha_a h_i + \beta_a s_i) + m_{ak} + \epsilon_{ak}.$$

Here $c_{ik}$ is the GC-MS-quantified concentration of compound $i$ in sample $k$, $\alpha_a$ and $\beta_a$ are the assay's sensitivities to the two mechanisms, $m_{ak}$ is the contribution from the matrix and probe artifacts, and $\epsilon_{ak}$ is noise. The mechanism sensitivities $\alpha_a,\beta_a$ are calibrated once on pure reference antioxidants.

Our second ingredient is a differential design. Each sample is measured both intact and after headspace stripping. Taking the difference cancels most of $m_{ak}$, and GC-MS of the stripped headspace quantifies exactly what was removed.

The claim is that the two ingredients together recover per-compound capacities, and flag genuine cross-assay inconsistencies, in cases where per-peak correlation and unstructured regression cannot.

The neutrino analogy supplies design logic, not a physical link. The combined reactor and accelerator analyses in Improved Constraints on Sterile Neutrino Mixing from Disappearance Searches in the MINOS, MINOS+, Daya Bay, and Bugey-3 Experiments and Limits on Active to Sterile Neutrino Oscillations from Disappearance Searches in the MINOS, Daya Bay, and Bugey-3 Experiments showed something neither experiment could show alone. The appearance amplitude factorizes as $4|U_{e4}|^2|U_{\mu4}|^2$. Each single-channel experiment bounds only one factor, so excluding the anomaly region required combining both channels under the shared parameters. That is the role our shared $h_i, s_i$ play across assays.

The same combined analysis also warns us about a weakness. It sums independent $\chi^2$ surfaces and assumes the systematic uncertainties of different experiments are uncorrelated. Different antioxidant assays share artifacts such as solvent and color interference, so we model $m_{ak}$ as correlated across assays rather than repeating that assumption.

The near/far idea comes from Search for Sterile Neutrinos in MINOS and MINOS+ Using a Two-Detector Fit. There, a shared covariance between detectors cancels flux and cross-section uncertainty. The stripped aliquot is our near detector.

The neutrino mass problem contributes one further, sobering lesson. Oscillations measure only mass differences, not absolute masses. Our differential design likewise identifies only the volatile-attributable part of activity, relative to a reference antioxidant. We do not claim absolute capacities.

Motivation from Existing Work

Three studies define the gap.

Chemical Composition, Antioxidant and Antimicrobial Activities of Essential Oil from the Leaves of Lindera fragrans Oliv. identifies 62 compounds covering 76.45% of the oil, dominated by spathulenol at 27.63%. It reports DPPH scavenging with a half-maximal inhibitory concentration of 33.01 mg/mL, against 23.33 mg/mL for the reference antioxidant Trolox. The same oil is inactive in the ABTS and β-carotene assays, and no compound is held responsible.

Physicochemical, Aroma Compounds, Microbial Community, and Antioxidant Capacity of Huangjiu-Based Functional Liqueur Fermented with Edible Herbs runs five assays alongside GC-MS and ion-mobility profiling. It links the two only by Pearson correlation.

In-Silico Antioxidant Activity and GC-MS Analysis of Root Extract of High-Altitude Himalayan Plant: Erigeron multiradiatus replaces measurement with molecular docking.

All three motivate the proposal: each shows that compound-level attribution is missing. None fits the assays jointly, and none controls for the matrix.

A fourth study constrains the model. Structural Elucidation and Antioxidant Activities of Proanthocyanidins from Chinese Bayberry (Myrica rubra Sieb. et Zucc.) Leaves documents steric hindrance: bulky molecules cannot reach the DPPH radical site, which depresses DPPH readings relative to FRAP. A purely linear two-mechanism model will therefore misfit bulky compounds. We allow a per-compound, per-assay accessibility factor, regularized toward 1, and we report it whenever the data demand it.

The method for turning raw GC-MS data into reliable concentrations $c_{ik}$ comes from two sources. Algorithmic Learning for Auto-deconvolution of GC-MS Data to Enable Molecular Networking within GNPS provides MSHub, which separates overlapping peaks using non-negative matrix factorization at about 1 minute per file, with 99.17% correlation to the reference tool XCMS. Feature-based Molecular Networking in the GNPS Analysis Environment provides feature alignment that keeps isomers apart. For unidentified peaks, Predicting Electron-Ionization Mass Spectrometry using Neural Networks offers candidate structures at 54.3% recall@1. Each peak still receives its own capacity whether or not it is identified, which matters given the 23.55% of the Lindera oil that was never assigned a structure.

Research Program

We would build and validate the attribution pipeline in three stages.

The first stage is calibration. We measure DPPH, ABTS, FRAP and ORAC on roughly twenty pure reference antioxidants spanning phenols, terpenoids and thiols. From these we fix the mechanism sensitivities $\alpha_a,\beta_a$ and the prior on correlated matrix terms.

The second stage is design. For each real sample we prepare a dilution series and a headspace-stripped twin, which gives many observations per sample instead of one. HS-SPME, a solvent-free method that pulls volatiles onto a coated fiber, then quantifies what stripping removed. The differential response $y_{ak}^{\text{intact}}-y_{ak}^{\text{stripped}}$ enters the joint fit with a shared covariance, mirroring the two-detector fit.

The third stage is inference. A Bayesian non-negative fit returns posterior capacities $h_i, s_i$ for every resolved peak. It also returns a goodness-of-fit tension score that flags any compound whose activity in one assay is not mirrored in the assays that share its mechanism. This is the analogue of an appearance signal that the disappearance channels do not support.

Baselines

The naive reference point is per-peak Pearson correlation between GC-MS peak area and each assay value across samples. This is the method of the Huangjiu study, run on our identical samples and peak table.

The strongest existing statistical competitor is partial least squares regression from the full peak table to each assay separately. It is the standard multivariate tool and is the most likely to match us. We give it the same deconvolved concentrations, the same samples including the dilution points, and cross-validated component selection with the same number of tuning trials as our prior hyperparameters.

The strongest experimental competitor is offline fraction collection: preparative separation into wells followed by plate-reader assays on each fraction.

No reported result was found on Lacuna for compound-level attribution accuracy by any of these methods. The only directly comparable reported number is the Lindera bulk DPPH result above, which is a whole-mixture endpoint.

Beating the baselines means two things on synthetic mixtures with known ground truth. First, lower rank error and lower absolute error in recovered per-compound Trolox-equivalent capacity. Second, correct identification of which compounds are mechanism-selective.

Two ablations isolate the ingredients. The first drops the shared $h_i,s_i$ structure and fits each assay independently. The second drops the stripped twin and fits intact samples only. If either ablation matches the full model, that ingredient is not earning its place.

Why not just…?

Why not just correlate peaks with assay values, as current studies do? With tens of samples and dozens of correlated peaks, correlation credits whichever compounds co-vary with the true active compound. Correlation also has no way to separate volatile from matrix contributions. The Pearson baseline on spiked mixtures tests this directly. If it recovers the planted actives as well as our model does, the proposal fails.

Why not just run partial least squares or another multivariate regression per assay? This is the real threat. With enough diverse samples, regression alone might recover attributions. The structured model can win only where assays disagree for mechanistic reasons, as in the Lindera DPPH-versus-ABTS split, or where the matrix contributes substantially. We therefore include mixtures designed to have exactly these properties, and mixtures without them. If partial least squares ties us on both kinds, we will report that the mechanism structure adds only interpretability.

Why not just collect fractions and assay them offline? Fractions merge co-eluting isomers; feature-based networking separates seven isomers at a single mass-to-charge value that classical clustering merges into one. Volatile terpenoids also evaporate or oxidize during handling. Fraction collection remains the best available ground-truth check on real samples, so we use it as a validator on a subset rather than dismissing it.

Why not just assay purified standards one by one? This works for identified, purchasable compounds. It misses unidentified peaks and any synergy or antagonism between compounds. We measure every planted mixture both as a mixture and as the sum of its pure standards. The size of the gap tells us whether this alternative is sufficient.

Proposed Experiments

The decisive test uses synthetic mixtures built from authentic standards of the Lindera constituents (spathulenol, ledol, β-caryophyllene, α-cadinol), together with planted phenolic and thiol actives of calibrated capacity. These are embedded in a non-volatile matrix of sugars, proteins and polyphenols that contributes known background activity. We generate about 60 mixtures with partially co-eluting pairs. The main metrics are Spearman rank correlation and mean absolute error of recovered capacities against ground truth, and the area under the curve for detecting mechanism-selective compounds.

Real-sample validation applies the pipeline to the Lindera oil and to Huangjiu variants. The question is whether the DPPH-only activity of Lindera concentrates on one or two peaks. That prediction is then checked by fraction collection and pure-standard assays.

The decisive outcomes are clear in advance. The proposal succeeds if the full model beats partial least squares on matrix-heavy and mechanism-split mixtures, and if both ablations degrade performance. It fails, and we will say so, if partial least squares matches the full model across all mixture classes.