Some of the most commonly sold milk products in Switzerland were purchased in stores from three main supermarket chains. Four liquid milk samples and two milk powders were collected and stored in laboratory at 6°C. Raw milk was first directly collected at two different dates into a glass flask (250 mL, particle certified, ThermoFisher Scientific) at a farm located in France (Jura country) using a milking machine (deLaval), then transported in a cold container to Switzerland and finally stored at 6°C in laboratory before analysis. Samples used for this study are listed in the table S2.
Milk sample digestion.
Powdered milk samples were initially reconstituted with ultrapure water in a glass flask (1000 mL; particle certified, Thermofisher Scientific) by diluting 25 g test portion in 175 g ultrapure water (Lichrosolv, Merck); 2 mL of multi-enzymatic detergent (Prozyme, Borer Chemie AG) were added in the reconstituted powdered milk and shaken for at least 15 minutes at 40°C into a water bath GFL-1083 Milian.
Processed and raw milk samples were transferred to glass flasks and covered with a glass watch. The samples were digested using a preparation that consists of 2 steps: enzymatic digestion and hot alkaline hydrolysis of milk samples, as described below.
25 mL liquid or reconstituted milk sample and 20 mL ultrapure water (Lichrosolv, Merck) were added into a glass flask (1000 mL; particle certified, ThermoFisher Scientific); 2 mL of multi-enzymatic detergent (Prozyme, Borer Chemie AG) were added and mixed for 2 minutes at 40°C in the container; 10 mL of chelating agent sodium ethylene diamine tetra acetate (EDTA-Na 0.5 M, pH 8, Invitrogen, ThermoFisher Scientific) were added and mixed for 3 minutes at 40°C in the container; 2 mL of alkaline solution tetramethyl ammonium hydroxide (TMAH 25% v/v Sigma-Aldrich) were added into the glass container that was finally immediately put in a microwave (Panasonic NN) at a power of 1000 watts for a maximum of one minute to reach about 80°C. Hot digested milk sample was then immediately submitted to filtration process.
Milk sample filtration.
Hot digested milk sample was directly poured into a glass funnel (100 mL, Sterlitech) mounted onto a filtration unit (glass holder with13 mm frit, stainless steel vacuum manifold, Rocker 400 Vacuum Pump, 220V/50Hz, Sterlitech). Contents of the funnel pass through a Si filter (Silicon filter, 10 x 10 mm size, 500 µm thickness, 5 µm porosity, SmartMembrane) under a vacuum for a maximum of 5 minutes. Home-made filter holder (stainless steel, 10 x 10 mm, 25mm), rubber hole seals (EPDM, 25 mm diameter, 2 mm thickness, 4 and 8 mm diameter holes), seal holder and cover (stainless steel, 25 mm, 8 mm diameter hole) were used for a restricted filtration area of ca 14 mm2 (i.e. 4 mm diameter area) on silicon filter (Figure S1). Retained digested material was successively flushed with first 5 mL water then 5 mL nitric acid 5% v/v and finally 10 mL water before being stored in a closed glass container before RM analysis.
Mitigating laboratory contamination.
A major challenge in the measurement of MPs is to avoid contamination of the test sample when handled in the broader laboratory environment, that includes for example particles from the ambient air, clothing, utensils, chemicals, and so forth. Meticulous study of the different possible steps that may introduce “foreign” MPs is a prerequisite to ensure reliable measurement. Therefore, emphasis on lowering the method blanks (i.e. blank following all the method process but omitting the sample matrix) besides the usual method validation parameters is pivotal. Class 2 biosafety cabinet, filtration unit, dedicated glassware and spectroscopy equipment (µRaman and optical microscopy) were installed in clean laboratory rooms with positive air pressure.
Control of laboratory environment was performed measuring MPs retained on some silicon filters put on laboratory benches and in biocabinet for one day before analysis. Grade air purifier (APS-500, Kynio) and particle counter (PC 220 Trotek) were used in the laboratory for mitigating and monitoring particles respectively. Alkaline solution (TMAH) is pre-filtered through 5 µm Ag filter (Sterlitech) whereas all other reagents used for milk preparation (i.e. multi-enzymatic detergent solution, EDTA-Na and ultrapure water) are pre-filtered through 0.65 µm PVDF filter (Merck) using stainless steel filter holders (Merck). Specific 1 L or 250 mL-Particle Free (PF) glass bottles (Thermo) are used for digestion procedure.
All the glass vessels (funnel, beakers, glass bottles), restrictor seals and filter holders were pre-cleaned thoroughly with ethanol 70% v/v and rinsed 3 times with ultrapure water (Merck).
The Si filter was pre-cleaned in a glass container with a pre-filtered aqueous solution of multi-enzymatic detergent (Prozyme, Borer Chemie AG) 10% v/v and finally rinsed with pre-filtered ultrapure water (Merck) in another glass flask. All the reagents used in filtration step (i.e. water, HNO3 (Merck) were pre-filtered on 0.6 µm PVDF filter (Merck)). Method blanks (i.e. blanks following all the method process but omitting the sample matrix) were systematically analyzed for each series of analysis.
The two quality criteria used for both efficient digestion and filtration steps were the digestion efficiency and recovery rates of spiked polymers in water and real sample.
Digestion efficiency was verified by counting remaining digested particles and the determination of total related surface covered in the filter by the digested matrix. The particle counting and determination of surface covered was performed using a digital optical microscope VHX-6000 (Keyence). A final maximum decision limit of 300 (with a maximum of 5% surface covered) and 600 retained particles per mm2 of filtration area (with a maximum of 30% surface covered) were established for procedural blank and digest sample, respectively. This criterion ensures that the sample was digested efficiently and that the filter was not overloaded with organic residues, which could potentially hide MPs or even clog the filter.
Raman method was used to determine the MPs recovery rate that was calculated by spiking one ultrapure water without alkaline digestion step and on one cow’s milk sample (Brand C − 0.1% fat) following the whole procedure of digestion. Different MP standards in solution (PMMA, PS, PA, PE and PP Sigma-Aldrich) were internally prepared and calibrated by spectral flow cytometry (FC, Cytek – Nestlé Research, Lausanne) in water containing sodium dodecyl sulfate detergent (SDS, Invitrogen) at 5% m/v. Polymer standard solutions used for recovery rates are listed in Table S3.
SEM micrographs of the Si filter surface combined with EDX analysis of the encountered particles were acquired for two spiked samples (ultrapure water and cow’s milk (Brand C − 0.1% Fat)) to identify and check the physico-chemical integrity of the MPs without and with alkaline digestion respectively.
SEM imaging - EDX analysis.
Simultaneous particle imaging and elemental chemical composition analysis were performed to check the quality of digestion of one spiked ultrapure water and one spiked cow milk sample (Brand C − 0.1% fat) employing a scanning electron microscope (SEM) “Quattro S” from Thermofischer equipped with an Energy Dispersive X-Ray Spectroscopy (EDX) detector “Xmax 50 mm2” from Oxford Instruments.
Prior to analysis, the Si filters were coated with either 5 nm of gold (for imaging purposes) or with 5 nm of carbon (for combined imaging and chemical composition).
Imaging of single particles was performed at an accelerating voltage of 5 kV in secondary electron and backscattered electron modes.
EDX spectra were acquired on single particles at 20 kV to confirm the organic or mineral nature of particles and subsequently at 5 kV to improve C, O and N detection. A minimum of 300000 counts were acquired per measurement.
Validation of Raman methodology.
Spiking experiment of one processed cow’s milk sample (Brand C – 0.1% fat) using Raman method was performed using different polymer standard solutions (PMMA, PS, PA, PE and PP Sigma –Aldrich) internally prepared and calibrated by flow cytometry (FC) in water containing sodium dodecyl sulfate detergent (SDS, Invitrogen) at 5% m/v with the polymer concentrations displayed in Table S3. Spiking tests were used not only to check the efficiency of digestion and method recovery, as well as possible changes in the Raman fingerprint spectrum. A spectral change could be used as indicator of chemical degradation of the MPs due to the alkaline digestion step.
A confocal µRaman Labram HR Evolution (Horiba, SAS France, Villeneuve d’Ascq, France) equipped with an EMCCD detector, 50x magnification long working distance objective (Olympus® ,NA = 0.5), dark field system and 532 nm solid-state laser (power 50 mW), acquisition time of 0.1 second in a spectral range of 550 cm− 1 to 2000 cm− 1 (resolution of 4 cm− 1) was used to acquire the Raman spectrum. This instrument setting generated a laser beam of ca 1 µm.
The detection and identification of microplastics on the surface of Si filter was performed by the point-by-point mapping approach using steps of 5 µm between two consecutive points. The distances of 5 µm between two consecutive points combined with the laser beam size of ca 1 µm allowed us to confirm that this approach can detect MPs with dimensions equal to or greater than 4 µm. On the other hand, the dimension of ca 1 µm of the laser beam also allowed us to confirm that this approach can potentially detect MPs smaller than 4 µm if the particle was close enough to where the laser beam was irradiated to collect the Raman spectrum.
In this approach, 2 areas representing 50% of total filter surface (i.e. 7 mm2) were consecutively analyzed (Figure S2). The instrument takes 2 times 10 hours to perform point-by-point mapping with step of 5 µm with a final acquisition of about 250,000 Raman spectra. A maximum limit of 600 retaining particles per mm2 of filtration area has been established for ensuring both correct µRaman mapping and imaging treatment using Labspec® 6.5 software.
Raman data processing.
After Raman spectra acquisition using LabSpec® 6.5 software, an in-house software was developed for identification and characterization of microparticles. The training data set was constructed using Raman spectra collected from exogenous MPs found in food samples and commercial MPs spike-in milk products. Such diverse set of MPs sources allowed a better representation of the possible range of signals for each given polymer. Furthermore, a wide range of signal quality was selected to better represent the noisier signals that can be found in theses samples.
The spectra database contained few spectra with slightly different spectral range. However, this small variation in the spectral range is not compatible with most machine learning algorithm, including random forest. Therefore, the spectra range and measurement points were refitted (559 cm− 1 to 1990 cm− 1, every 3 cm − 1) using R stats splinefun function to allow consistent measurement points of the spectra. Baseline correction was performed using the R package “baseline” to remove the background noise from the spectrum (61). A set of “universal” parameters was chosen to best fit the observed cases. To reduce heavy fluctuation of the signal, the spectra were smoothed using another spline function included in the R package hyperSpec (62). Parameters were chosen so that all data points were used with a reduced smoothing parameter to avoid any degradation of the signal. Finally, the spectra values were scaled over their respective standard deviation.
The Random Forest machine learning algorithm was fed with the preprocessed data. The algorithm was trained with the training set with of 1000 estimators and will classify each spectrum in one of the MP class or in the non-MP (NMP) class (63).
Using the physical position of each measurement, the neighboring signals of the same class of polymer were grouped into particles. Using the R package “lattice” (64), a graph represented the physical connections between spectra. The edges of the graph were then kept if the maximum distance between two spectra of a same polymer class was two. Effectively, this allowed clustering of polymer signals based on their relative physical location, while jumping over “missed” signals. Missed signal could be caused by a short desynchronization between the computer clock and the speed of acquisition of the spectrum.