Ionization Efficiency Prediction Models for Semi-Quantitative Mass Spectrometry Screening

Spray

Electrospray ionization converts liquid eluent into charged gas-phase analytes through droplet fission, charge concentration, and solvent evaporation. In untargeted screening of plastics migrates, a high-resolution mass spectrometer records signal intensities as peak areas, but equal peak areas for two different chemicals do not mean equal mass. The ionization process is extremely selective.

Non-polar paraffinic hydrocarbons, sterically hindered cyclic esters, and quaternary ammonium slip additives elute with physical response factors that can differ by up to six orders of magnitude under identical mobile phase conditions. Assigning a universal surrogate response factor to every uncalibrated peak leads to false passes on toxic migrants and unnecessary lot quarantines for benign processing aids.

Liquid chromatography coupled with electrospray ionization creates a practical problem during packaging characterization. Regulation (EU) 10/2011 mandates the identification and toxicological evaluation of non-intentionally added substances that migrate into food simulants. When extracting a multilayer polyethylene barrier film with ninety-five percent ethanol for ten days at sixty degrees Celsius, hundreds of chromatographic features appear across the total ion chromatogram.

Authenticated reference standards exist for fewer than fifteen percent of these features. Synthesizing obscure photoinitiator fragments, adhesive degradation products, and oxidized antioxidant oligomers takes weeks of bench work. Laboratories often quantify uncalibrated peaks by comparing their areas directly against a single internal standard, commonly bisphenol A-d16 or 2-ethylhexyl 4-dimethylaminobenzoate.

That shortcut introduces quantitative errors exceeding a factor of one thousand.

A single internal standard applied across a full chromatographic run yields concentration errors spanning three orders of magnitude in positive electrospray mode.

Charge transfer during desolvation depends on solution chemistry, gas-phase thermochemistry, and interface mechanics. Basic analytes with high gas-phase proton affinities readily strip protons from hydronium ions in positive electrospray mode, producing strong precursor ion signals even at trace concentrations. Acidic compounds with high gas-phase acidity drop protons to solvent molecules in negative mode.

Analytes without ionizable functional groups rely on competitive adduction with background sodium, potassium, or ammonium ions. When an eluting peak co-elutes with non-volatile plasticizer residues, ion suppression steals charge from trace migrants. Physical ionization efficiency measures this conversion as the ratio of gas-phase ions reaching the detector to the total molar amount of neutral analyte entering the electrospray needle.

Untargeted mass spectrometry screening without individual reference compounds is often assumed to provide a reliable ceiling on non-intentionally added substance migration.

Covariate

Molecular properties dictate whether an organic compound gains or loses charge inside a charged solvent aerosol. Over the past decade, predictive chemometrics moved from empirical corrections toward trained in silico quantitative structure-property relationship models. These systems map multidimensional molecular descriptors directly to measured relative ionization efficiencies.

By calculating structural features from two-dimensional chemical connectivity tables, predictive models generate compound-specific response factors without requiring an authentic physical standard at the bench.

Physicochemical molecular descriptors retain physical meaning across different chromatographic regimes. The octanol-water partition coefficient governs how an analyte partitions between an evaporating droplet’s core and its charged surface. Hydrophobic molecules move quickly to the air-water interface, favoring field desorption into the gas phase.

Topological polar surface area measures the electrostatic footprint that hinders ion transfer across the boundary layer, while dynamic gas-phase basicity defines the energy released when a neutral molecule accepts a proton. Combined with retention parameters, these descriptors capture the solvent composition at the exact moment the analyte reaches the emitter nozzle, accounting for shifts as the gradient transitions from aqueous buffer to organic modifier.

Machine learning regression models correlate structural descriptor matrices with log-transformed experimental ionization efficiencies. Gradient boosted decision trees, random forests, and deep feed-forward neural networks achieve cross-validated root-mean-square errors between 0.38 and 0.65 log units across diverse chemical libraries. An error of 0.50 log units corresponds to an average prediction uncertainty within a factor of 3.16 in linear concentration space ~ narrowing the traditional four-order-of-magnitude semi-quantitative uncertainty down to a workable range.

Deploying predictive ionization models requires a screening setup that maintains chromatographic stability over long sequences. Three operating factors dictate how well models transfer between different mass spectrometers:

• Source Geometry Normalization eliminates signal bias caused by varying emitter-to-inlet distances and counter-electrode angles across different instrument designs.
• Solvent Composition Tracking recalibrates predicted responses to match the exact fraction of organic modifier at the analyte’s retention time.
• Eluent Additive Consistency controls proton availability by keeping formic acid, ammonium acetate, or ammonium fluoride at fixed ionic strengths.

Prediction accuracy drops when an unknown compound falls outside the applicability domain of the training set. Organofluorine processing aids, polybrominated flame retardants, and complex metal-organic pigments carry elemental distributions poorly represented in standard small-molecule libraries. Defining the exact boundaries where a non-intentionally added substance can be predicted without custom retraining remains an open question in computational chemistry.

Threshold

Analytical measurements on food-contact migrates connect directly to statutory toxicological thresholds. Article 19 of Regulation (EU) 10/2011 requires packaging manufacturers to assess non-intentionally added substances using internationally recognized risk assessment principles. When toxicological data on an unlisted degradation compound is unavailable, safety dossiers rely on the Threshold of Toxicological Concern framework.

Under European Food Safety Authority guidance, an uncharacterized migrant without structural alerts for genotoxicity is evaluated against the Cramer Class III threshold of ninety micrograms per person per day. Based on standard European consumption assumptions of one kilogram of packaged food per day, that threshold translates to a migration limit of ten micrograms per kilogram of food simulant.

How ESI Response Disparities Skew Toxicological Triage?

Untargeted screening filters detected peaks against that ten microgram per kilogram boundary. When migration concentrations are calculated using uncorrected single-surrogate calibration, a compound with poor electrospray response can produce a peak area well below the ten microgram cutoff even if its actual concentration in the simulant reaches eighty micrograms per kilogram. The packaging material receives a passing report and enters the retail market.

Subsequent audits by national surveillance authorities then identify the unassessed substance during confirmatory testing, triggering mandatory product recalls under Article 14 of Regulation (EC) 178/2002.

A substance whose actual migration exceeds ten micrograms per kilogram escapes regulatory review whenever poor ionization efficiency artificially suppresses its analytical signal.

Incorporating ionization efficiency prediction models changes how screening limits are set. Because machine learning models carry an uncertainty band, laboratories establish a conservative lower threshold. A typical analytical sequence to prevent false negatives runs as follows:

1. Calculate the predicted ionization efficiency and its ninety-fifth percentile lower prediction interval using trained descriptor models.
2. Set the conservative screening limit by dividing the toxicological threshold of concern by the model’s upper uncertainty factor.
3. Flag every peak exceeding this adjusted limit for structural identification using high-resolution tandem mass spectrometry libraries.
4. Procure or synthesize certified reference standards for flagged substances that fall into structural classes associated with mutagenic or endocrine-disrupting properties.

Prediction confidence depends directly on structural similarity between unknown migrants and the molecules in the training set.

Arithmetic

Evaluating migration from laminated barrier packaging shows the quantitative shift that occurs when moving from surrogate calibration to model-based response prediction. Consider a flexible retort pouch built from a twelve-micrometer polyethylene terephthalate outer film, a nine-micrometer aluminum foil core, and a seventy-micrometer cast polypropylene sealant layer bonded with a solvent-based polyurethane adhesive. Migration testing targets fatty food contact using European food simulant D1 (fifty percent aqueous ethanol by volume).

Exposure runs for ten days at forty degrees Celsius in a double-sided migration cell with a surface area to food volume ratio of six square decimeters per kilogram of simulant.

Analyzing the simulant extract by high-resolution liquid chromatography quadrupole time-of-flight mass spectrometry in positive electrospray mode reveals an uncalibrated peak eluting at 8.42 minutes with an accurate mass of 248.1645 daltons. High-resolution tandem mass spectrometry identifies the compound as 1,4-diazabicyclo octane dibutyl ether, a breakdown product from thermal cleavage of the polyurethane curing catalyst. Peak area for this feature measures 4,200,000 counts.

An internal standard of deuterated bisphenol A eluting at 7.15 minutes yields 5,500,000 counts for a known concentration of 20 micrograms per kilogram.

Conventional single-surrogate semi-quantification calculates migrant concentration through direct proportionality:

Concentration = (Peak Area of Unknown / Peak Area of Surrogate) x Concentration of Surrogate

Inserting the experimental instrument values produces an initial concentration estimate:

Concentration = (4,200,000 / 5,500,000) x 20 µg/kg = 15.27 µg/kg

This raw surrogate calculation yields a borderline concentration slightly above the ten microgram per kilogram limit, which might prompt a packaging team to consider reformulating an entire adhesive line.

Predicting the ionization efficiency corrects this baseline. Extracting twenty-four molecular descriptors from the identified structure yields a log octanol-water partition coefficient of 2.14, a topological polar surface area of 18.7 square angstroms, and a calculated gas-phase basicity of 945 kilojoules per mole. The gradient boosted decision tree model predicts a log relative ionization efficiency of plus 1.42 relative to deuterated bisphenol A ~ meaning the breakdown product ionizes 26.3 times more efficiently than the surrogate standard under these gradient conditions.

Correcting the concentration calculation with the predicted response factor inverts the operational decision:

Corrected Concentration = Uncorrected Concentration / Predicted Relative Ionization Efficiency

Corrected Concentration = 15.27 µg/kg / 26.3 = 0.58 µg/kg

The actual migration is 0.58 micrograms per kilogram ~ well below the ten microgram limit ~ saving the facility from an unnecessary four-hundred-thousand-euro adhesive line shutdown.

The reverse scenario shows the risk of false negatives. Consider the oxidized antioxidant byproduct tris(2,4-di-tert-butylphenyl)phosphate. Its large size, lack of basic amine groups, and low droplet-surface affinity suppress ionization, giving it a predicted relative ionization efficiency of 0.08 compared to the surrogate standard.

A peak area that yields a nominal surrogate reading of 4.2 micrograms per kilogram actually corresponds to 52.5 micrograms per kilogram in the simulant. Relying on uncorrected surrogate screening would pass a material migrating toxic degradation products at five times the legal limit.

Relying on uncorrected surrogate peak areas leads to flawed safety filings, unnecessary line shutdowns, and unexpected detentions when enforcement laboratories perform calibrated confirmatory testing.

Warrant

Declarations of conformity protect buyers only when the supporting technical dossier stands up to regulatory scrutiny. Article 16 of Regulation (EC) 1935/2004 requires operators to maintain documentation proving compliance with safety thresholds. When an enforcement agency or auditor requests the file for a plastic food contact material, the declaration sheet alone offers no legal protection without solid analytical data.

If the screening behind it relies on uncorrected response factors, the compliance argument collapses.

Auditors inspecting compliance files look for four key elements in untargeted screening reports. First, the documentation must state the calculations used to set concentration limits for unidentified peaks. Second, it must specify the calibration standards used and the uncertainty factor applied to the screening boundary.

Third, it must show that uncalibrated peaks fell within structural applicability domains before predicted values were accepted. Fourth, the file must contain complete instrument logs covering electrospray parameters, spray voltage, gas temperatures, and mobile phase additive lot numbers. Missing any of these details leaves the dossier vulnerable during an audit.

A migration screening report that quotes non-detected results against a single internal standard fails regulatory scrutiny during a competent authority audit.

Commercial supply agreements increasingly use specific warranty language to cover non-intentionally added substance verification. Sourcing contracts that simply mandate compliance with Framework Regulation (EC) 1935/2004 often fail to protect downstream buyers if testing relies on uncorrected semi-quantitative screening. Modern procurement contracts address this by including explicit analytical requirements in technical schedules:

The supplier warrants that all non-intentionally added substance migration screening documentation provided in support of the Declaration of Conformity applies compound-specific ionization efficiency predictions with documented ninety-five percent prediction intervals, ensuring no unassessed migrant exceeds the applicable Threshold of Toxicological Concern within a five-fold analytical uncertainty envelope.