Secure Food-Allergen Determination by Combining Smartphone-Based Raw Image Analyses and Liquid Chromatography–Mass Spectrometry for the Quantification of Proteins Contained in Lateral Flow Assays

Abstract

The current food safety testing system, based on laboratory-based quantification, is difficult to scale up in line with the growth in the export market and does not enable traceability through the nodes of the food supply system. Screening assays, for example, lateral flow assays (LFAs), can improve traceability but often lack the required reliability to guarantee compliance. Here, we present an alternative pipeline for secure on-site compliance testing, using allergens as a case study. The pipeline features smartphone-driven LFA quantification and an liquid chromatography–mass spectrometry (LC–MS) method enabling direct quantification of the allergens contained in the LFA. The system enables swift and objective screening and provides a control measure to verify LFA assay reliability. For the smartphone assay, 8-bit RGB and grayscale colorimetric channels were compared with 16-bit raw intensity values. The latter outperformed RGB and grayscale channels in sensitivity, repeatability, and precision, while ratiometric ambient light correction resulted in excellent robustness for light-intensity variation. Calibration curves for peanut determination using two commercial LFAs featured excellent analytical parameters (R² = 0.97–0.99; RSD 7–1%; LOD 3–7 ppm). Gluten determination with a third commercial LFA was equally established. A prediction error of 13 ± 11% was achieved for the best performing assay. Good performance–calibration curves (R² = 0.93–0.99) and CVs (<15%)– were observed for the analyte quantification from the LFA by LC–MS. The LOD for the LC–MS assay was 0.5 ppm, well below the LODs reported for the LFAs. This method creates a digital, fast, and secure food safety compliance testing paradigm that can benefit the industry and consumer alike.

The detection, quantification, and mitigation of contamination are essential to ensure food safety. Allergens pose a significant health burden to the global population, with 3.5–4% being affected by food allergies.¹ Allergic persons must follow avoidance diets to evade allergic reactions. This requires food products to be properly tested and labeled for the presence of known allergens. In the European Union, the United States, and Australia, only foods that have been intentionally incorporated into food products must be listed as ingredients. As such, trace amounts of allergens that have unintentionally entered the product (e.g., through cross-contamination) do not need to be declared. However, to protect consumers from allergic reactions and themselves from legal consequences, many food manufacturers voluntarily employ precautionary allergen labeling (PAL) systems. Unfortunately, as PAL systems are unregulated and largely overused, many consumers choose to ignore these warnings, which can lead to dire consequences including life-threatening anaphylaxis.^2,3 The adoption of quantitative threshold levels for reporting allergen content can improve the utility and effectiveness of food labels for the allergic population. The Voluntary Incidental Trace Allergen Labeling (VITAL) program (https://vital.allergenbureau.net/) is a good example of such a system and, if endorsed, this can considerably improve food safety.⁴ To implement such a system, quantitative, accurate, and reliable allergen tests are required on-site. However, current on-site detection methods are mainly qualitative while lab-based ELISA assays for food allergen detection often suffer from cross-reactivity and poor recovery issues. This makes the implementation of quantitative labeling difficult to implement and enforce.^4,5

Targeted proteomic approaches show great promise for laboratory-based quantification of allergens, because such methods do not suffer from cross-reactivity issues, have excellent multiplexing potential, and can be used as confirmatory analyses to screening tests, which further improves food safety.⁶⁻⁸ However, liquid chromatography–mass spectrometry (LC–MS) approaches are time consuming and performed on raw materials in a laboratory, often far from the production line where food is processed. Geballa-Koukoula et al., developed an ambient mass spectrometry approach for small-molecule detection from lateral flow assay (LFA) extracts.²³ This method is more rapid than conventional LC–MS and enables the confirmation of LFA screening results, albeit qualitatively, constituting a step forward for improved food contaminant analyses. On-site testing with a paper-based device such as a LFA can enable food manufacturers to test individual batches for cross-contamination. LFAs can empower consumers to monitor the safety of their own food as well and democratize analytical chemistry protocol execution.^9,10 LFAs are primarily qualitative, reporting the presence or absence of a target analyte.¹¹⁻¹³ However, LFAs are not as accurate or sensitive as LC–MS quantification. Moreover, LFAs can be misused leading to the reporting of erroneous results. A crucial limitation of LFA interpretation is the reliance on naked-eye readings. Variable environmental conditions, such as low levels of ambient light and glare, on top of perceptual differences make these assays susceptible to misinterpretation.¹⁴ This is compounded if a user has impaired color perception or is longsighted.¹⁵ Digital readers for the interpretation of LFAs can overcome these issues. However, such systems only work for LFAs of the specific manufacturer and are quite costly. Smartphone-based quantification is appealing because smartphones remove costs associated with additional hardware. Moreover, smartphones have the capacity to enable real-time geotagging of test locations, secure data storage and management via the cloud, and allow rapid communication to stakeholders.^11,12,16 However, smartphone-based LFA quantification has proved to be challenging with model-dependent spectral sensitivity and large inter-phone variations observed.^12,17 Previous studies have explored many different color channels, resulting in a plethora of colorimetric methods.^12,16,18 Usually, a camera’s image sensor is covered with a patterned light filter allowing only the red, blue, or green section of the visible light to reach a single pixel. The filter is often arranged in a Bayer pattern.¹⁷ The raw images contain pixel-wise light intensity information (often in a 12–16 bit scale) for red, green, or blue filtered light in a mosaic Bayer pattern. This minimally processed raw image is transformed through an internal image processing pipeline into a compressed RGB color image by interpolating “green, red, and blue” color values from neighboring pixels into each pixel and rescaling light intensity to 8 bit for each pixel. This process results in a loss of information, introduces uncertainty, and means that JPEG RGB values cannot be directly related to incident light.^17,19−21 As smartphones are evolving, it is now possible to retain the raw (DNG) 16-bit images on new smartphones. These images are uncompressed and have a much broader (65,536-step) dynamic range as an 8-bit (256-step) system. This may improve the performance of LFA quantification with a smartphone. Moreover, ratiometric corrections for ambient light variation may be more efficient for this approach due to the limited processing performed. Data analyses of these larger files may be cumbersome on older phones but is not expected to be inhibiting on novel models since processing speed of novel smartphones is approaching—and on occasions, exceeding—the speed of current-generation laptops.²²

The present study evaluates the opportunity of using DNG smartphone images and ratiometric ambient light correction for color quantification of LFAs via a smartphone followed by subsequent LC–MS analyses directly on extracts obtained from the LFAs. The performance of RGB color channels and gray scale from JPEG images were compared with raw intensity values from DNG images under various lighting conditions. The developed system was used for the quantification of three commercial allergen LFA test strips, one for gluten and two for peanut allergen. The LC–MS approach enables the direct validation of the LFA results by quantifying allergen specific peptides directly from the LFA. Opposed to completing an LC–MS assay on a separate sample, this novel approach enables the validation of any quantitative readings made from a mobile device and mitigates against sample management issues. Sample credentials can be captured by the smartphone linking the LFA identity with the sample time and placed while the sample and LFA content can be verified via LC–MS. Overall, a new evidence-based protocol is presented that can enable quicker and more cost-efficient compliance certification for food contaminants by fusing (i) a novel data analyses approach combining the use of unprocessed raw sensor image data with ratiometric ambient light correction resulting in improved robustness toward ambient light variation, and an increase in sensitivity and dynamic range with (ii) an LC–MS-method enabling the verification of both the allergen content and the performance of the screening assay by extracting and quantifying proteins from the LFA directly, instead of only verifying the allergen content found in the sample (Figure S1).

Combined, these methods constitute a highly novel pipeline that allows on-site quantification of allergens by image analyses, data transmission and automated auditing of the sample content, the LFA test performance, and the sampling procedure by subsequent LC–MS analyses of the LFA protein content. The pipeline has the potential to create significant impact in the food industry where food recalls due to unintentional cross contamination of allergen free food with allergens being a considerable issue causing health burden as well as economic cost due to brand damage and recalls. Moreover, the pipeline has excellent potential to be implemented for the quantification/control of other LFA in a variety of fields.

Materials and Methods

Materials

A 99.9% pure commercial peanut butter (from roasted Australian peanuts) was defatted and used as a standard for peanut allergen following previous work.⁸ Grain samples from the wheat cultivar Chara were obtained from the Australian pasture collection and Australian Winter Cereals Collection (Tamworth, Australia). The samples were manually inspected to exclude any foreign seed contaminations. Flour samples were obtained by milling the grains with a Metefem Hungarian Mill (model FQD2000, Hungary). Commercial allergen LFA test strips were acquired from 3M (Gluten Protein Rapid Kit and Peanut Protein Rapid Kit) and Romer (Peanut AgraStrip). Dithiothreitol, ammonium bicarbonate, formic acid, and phosphate-buffered saline tablets were bought from Sigma-Aldrich. LC–MS grade acetonitrile was purchased from Merck Millipore.

GNP Synthesis and Ambient Light Correction

GNPs were synthesized following the Turkevich method. Briefly, 500 μL of 100 mM HAuCl₄ was added to 194.5 mL H₂O and brought to boil in an Erlenmeyer flask with a condenser under stirring.²⁴ Sodium citrate solution [5 mL, 1% (w/v)] was then added, and the solution was boiled for 30 min while being stirred and then cooled gradually. This formed the stock solution which was stored at 4 °C. Nanoparticle size was estimated as 16 nm from UV measurements following,²⁵ and a 30× concentration was obtained by centrifuging (13,000 RCF; 30′; 20 °C). The 30× GNP stock was used to construct a calibration curve from a 2× dilution series from 30 to 0.9375 relative concentration units. Ten μL (n = 3) of each concentration was dropped on Whatman filter paper backed with two additional filter paper layers to ensure good pull through/absorption and dried. Spots were photographed with a Galaxy S21 Ultra (Samsung) as both RAW and JPEG images using 1/30, 1/45, and 1/60 shutter times and backlit conditions (Figure S2).

Image Processing

All photographs were backlit to reduce illumination variation. Both JPG (processed 8-bit RGB files) and DNG (raw, unprocessed, Bayer filtered 16-bit files) were recorded for each image taken. ImageJ (V1.53k) was used to extract R, G, and B channel as well as weighted and average grayscale values from the JPEG files. The same program was used to extract gray-scale intensity values from raw DNG files. The Samsung camera was operated in Pro-mode. The iPhone (model S6) was operated using the Pro-Cam application. The white balance was fixed at 5000 and ISO at 100, while shutter times were varied in optimization experiments with GNP spots and then fixed at 1/60. Figure S3 shows exemplary JPG and DNG images taken with the Android and iPhone from commercial LFAs for gluten and peanut detection using these fixed camera settings. Channel intensity values were averaged per drawn box using approximately 10,000 pixels per box. For RAW intensity readings, the DCRaw plugin was used to linearize images using the “totally RAW” document mode with a 16-bit output applying no interpolation or color, using the fixed camera white balance without automatically brightening the image and preserving the pixel orientation and aspect ratio.²⁶ RGB intensity values for each color channel of the accompanying JPEG image as well as the average and weighed grayscale values of the accompanying JPEG images were equally recorded. A data processing diagram showing the analysis procedure in ImageJ is shown in Figure S4.

Liquid Chromatography Multiple Reaction Monitoring Mass Spectrometry

An existing liquid chromatography multiple reaction monitoring mass spectrometry (LC-MRM-MS) method was followed for gluten quantification.²⁷ For peanut allergen analyses, an MRM method was developed herein. Briefly, FASTA sequences for the peanut allergens Ara h1, h2, h3, and h6 were imported into Skyline (v21.1.0.146, University of Washington)²⁸ and in silico digested with trypsin allowing no mis-cleavage, no modifications, and a minimum peptide length of 6 amino acids. The Prosit deep-learning algorithm was used to predict the top five monoisotopic product ions (1+ or 2+) for both y and b ions from the 2+ and 3+ precursors, within an ion match tolerance of 0.5 m/z.²⁹ Normalized collision energy was set at 33 (this was empirically determined to best match SCIEX instrumentation). Optimum collision energy values were predicted using the build-in collision energy equations within Analyst software (v1.7.1). Retention times (RTs) were predicted with Skyline using the experimentally determined RT times of 11 Biognosys iRT peptides and experimentally verified using a 60 s detection window in initial experiments. Three-co-eluting transitions were kept for use in the final LC–MRM–MS method which had a 40 s scheduled detection window and a 0.5 s cycle time. Skyline was used to integrate MS² ion intensity peaks. In optimization experiments, peak areas of the three most intense transitions were summed. For the final quantitative method, the ion ratio (IR) of the qualifying to quantifying co-eluting transition and the RT were used for identification. An accepted variance of 20% if IR > 0.5, 25% if 0.2 < IR < 0.5, and 30% if 0.1 < IR < 0.2 was used. A signal to noise ratio (S/N) > 3 was used for the limit of detection (LOD) and S/N ion >10 for the limit of quantification (LOQ).

LFAs and Extractions

Peanut calibration curves were constructed by serial dilutions from 2000 to 0.49 μg mL^–1 in phosphate-buffered saline (PBS; pH 7.2). For gluten, a calibration curve was constructed by serial dilutions from 8000 to 0.98 μg mL^–1 in PBS. The manufacturers’ LFA assay protocols were followed for all experiments. LFAs were photographed at the time indicated as suitable by the manufacturer. The sample, wick pad, and test line of each assay were carefully sliced from the nitrocellulose membrane immediately after being photographed and put onto 0.45 μM spin filters and stored at -80°C. Subsequently, 400 μL of extraction solution (8 M urea, 50 mM dithiothreitol) was added and spin columns were sonicated for 10 min followed by 30 min of shaking at room temperature and centrifugation (19,600g; 5 min). The supernatant was transferred to 10 kDa Amicon filters and spun for 15 min at 19,600g. Each column was washed thrice using a solution of 200 μL of 50 mM ammonium bicarbonate and 1 mM calcium chloride solution (ABC) (19,600g; 10 min). Trypsin (200 μL at 5 μg ml^–1 in ABC) was added, and extracts were digested while shaking for 1 h at 37 °C. Next, filters were spun (15 min; 19,600g), and the flowthrough was collected. Then, 200 μL ABC was added and filters were centrifuged (15 min; 19,600g). The flowthrough was pooled and evaporated to dryness in a speed vac. Lyophilized peptides were solubilized in 50 μL of acidified water spiked with iRT (0.1% FA, 0.05 iRT pmol μL^–1) and transferred to LC vials for LC–MRM–MS analyses.

LC-MRM-MS Analyses

Peptides were chromatographed (8 μL) with a Shimadzu Nexera UHPLC system with a Kinetex C18 column (100 × 2.1 mm; 1.7 μM particle size; 100 Å pore size). Solvent A was MQ with 0.1% FA. Solvent B was 90% acetonitrile, 0.1% FA. Flowrate was 0.4 mL min^–1. The chromatographic conditions (in % B) were isocratic (5%) for the first 0.2 min and then increased to 45% over 10 min, before being increased to 80% over 0.8 min. Conditions were then isocratic for 1 min. The % B was then decreased to 5% in 0.1 min and kept isocratic for 2.9 min. Eluted peptides were analyzed with a 6500 QTRAP (SCIEX) following.⁸

Software and Statistics

ImageJ V1.53k was used to extract RGB and RAW intensity values. Skyline v21.1.0.278, GraphPad 9.3.0, Biorender.com, and Inkscape 1.1.1 were used to prepare and assemble statistical analyses and figures.

Results and Discussion

Channel Choice and Ambient Light Correction

A calibration curve of GNP concentrations (n = 3) on filter paper was photographed (shutter times of 1/30, 1/45 and 1/60) with the Samsung in RAW and JPEG format to optimize the camera settings and test the performance of the ambient light correction method. R, G, B, average grayscale, and weighted grayscale measurements were extracted from 8-bit JPEG images, and intensity values were extracted from 16-bit RAW images. These values were fitted using a four-parameter log(dose)–response curve (Figure 1). The summary statistics (R²; LOD and IC₅₀) of the curve fits are shown in Table S1. Curves using ratio corrected values were mainly shown to have similar or slightly lower LOD and IC₅₀ values as curves using uncorrected measurements. This may be caused by a reduction in background illumination noise. The red color channel underperformed. This may be because varying red background noise disturbed the measurements, whereas grayscale or blue channel values were less affected by this artifact in the images. Ratio correction did have a slight negative effect on the fit (R²) of the calibration curve to the data across JPEG and RAW image formats (Table S1). This negative effect was the least prominent in the RAW ratio corrected images. A very notable finding is that ratio correction mitigates the impact of shutter time, as a proxy for light intensity, on the yielded intensity values (Figure 1C–H) in comparison with the uncorrected calibration curves (Figure 1A,B). Although this was observed to varying extents across the color channels within JPEG images, this effect was most distinct in the ratio corrected RAW images (Figure 1H). This suggests that raw images enable high quality corrections for light intensity changes within this straightforward ratiometric approach. In addition to this, LODs and IC₅₀ values were comparable or lower as the RGB and grayscale LODs and IC₅₀ values for all trialed shutter times, while the R² values were superior. The 1/60 shutter time setting produced comparable R² values and better sensitivity across the board. This is likely due to lower light exposure, lowering the impact of environmental light conditions; this shutter time was chosen from here on. Additionally, only ratio metric values were used from here on since this method showed excellent robustness against ambient light variations, which are unavoidable in the field.

Figure 1.

Calibration curves from processed images of GNP infused filter paper. (A,B) Plots using uncorrected intensity values of JPEG color channels (red, blue, and green), average and weighed grayscale (A) and raw intensity values (B). (C–G) Plots using ratio (signal/white background) corrected JPEG color and grayscale channels. (H) Plots using ratio corrected raw intensity values. All images (n = 3) were taken with a Samsung Android in pro-mode with fixed iso, white balance, and zoom settings. Shutter times were varied (1/30, 1/45, and 1/60) as indicated in legend.

Raw Format Performance across Mobile Devices

Interphone variation has been identified as a significant issue for LFA quantification.¹² The ratiometric approach improved calibration curve overlap compared to uncorrected calibration curves (not shown), but this did not completely resolve the issue (Figure 2A–F). Furthermore, clear differences in the amount of calibration curve overlap and shape can be observed for the tested image and color/intensity formats applied. Red and blue channel JPEG calibration curves enabled the closest fits between the two phone models. Notably, ratio corrections decreased the R² of the Samsung calibration curves but improved or did not affect the fit of the iPhone calibration curves compared to the R² values obtained for the uncorrected data in the JPEG image format (Table S2). Thus, ratio correction varied in effectiveness depending on the phone model when the JPEG format was used. For the raw images, R² values were excellent for both phone models and did not differ between uncorrected and ratio-corrected calibration curves. The raw intensity values also led to particularly low relative standard deviations (RSDs) for both phones (iPhone 1.05 ± 0.51%; Samsung 1.41 ±0.42%); and was the only measurement type that enabled the detection of GNPs at a concentration below a relative value of one; and was the only measurement type that showed no plateau in the higher concentration range of the GNPs for both phone models. As such ratio corrected values were used from here on as these were considered to outperform all JPEG color channel and intensity measurements in terms of ambient-light variation robustness, inter-phone variation, dynamic range, precision, and sensitivity.

Figure 2.

Calibration curves of GNP concentration and intensity in color channels of JPEG and RAW images captured using Android (black) and iPhone (red) mobile phone models at 1/60 shutter time using red (A), green (B), blue (C), average gray scale (D), weighed gray scale (E), and raw channel intensity values. Validation of image quantification using allergen LFA test strips.

Calibration curves were constructed using raw intensity values for two commercial peanut LFAs (Romer and 3M) and a gluten LFA (3M) using both phone models (Figure 3; Table 1). The iPhone underperformed compared to the Samsung in terms of R² values and the Δ ratio observed between the maximum and minimum signal across all investigated assays. Nonetheless, high R² values (0.94 < R² < 1.00) and exceptionally low repeatability RSDs (between 1.0 and 7.0%) were obtained for all calibration curves (Table 1). Overall, the LODs and linear ranges for the iPhone slightly outperformed (maximum ∼2-fold improvement) the Samsung values. Sensitivity of the peanut assays varied between 3.4 (iPhone) and 6.8 (Samsung) ppm. These values are above the reported LODs of these assays by the manufacturer (1 ppm) but remain in the same order of magnitude. Notably, no clear test-line coloration was observed upon visual inspection of the test lines at peanut levels below ∼5 ppm for both peanut assays confirming that visual and smartphone-derived LODs were similar for these assays for both tested models. The visual (∼500 ppm) and smartphone-derived (688.6 ppm for Android; 295.6 ppm for iPhone) LODs for the 3M gluten assay were considerably higher than reported by the manufacturer as being detectable by the naked eye (∼5 ppm). The reasons for this discrepancy are unclear but may be related to matrix effects. Additionally, the previously used ratiometric correction, whereby test line intensity values were divided by the white background intensity values, did not produce acceptable results for the gluten assay. Instead, the test-line intensity values were divided by the control line intensity values here to obtain acceptable calibration curves. This indicates that the accuracy of different ratiometric correction techniques may be dependent of assay specific parameters such as run time, membrane porosity, and antibody specificity—influencing random control line and blank area intensity variation. Thus, the denominator for this correction needs to be optimized depending on the assay. Strikingly, the iPhone calibration curves of the Romer and 3M assays show excellent overlap, while the Android calibration curves do not (Figure 3A). This indicates that assay specific calibration may not be necessary to correct for commercial assay variation for certain phone models with this approach. Prediction accuracy was tested for the peanut LFA assays with new samples tested one month after the calibration curve generation. This test was performed at 12 and 4 ppm for the Romer assay and 10 and 5 ppm for the 3M assay. The average error on the predictions was 38 and 37% for the iPhone Romer and 3M assays, respectively. The average error on the predictions was 13 and 47% for the Samsung Romer and 3M assays, respectively. The quantification of the Romer assay with the Samsung produced excellent results. Moreover, the assays enable peanut determination over a range of more than two orders of magnitude. As such, even a 50% (or 2-fold) prediction error may still be acceptable, especially as a semi-quantitative result that can be followed up with LC–MS based quantification.

Figure 3.

Smartphone based quantification of commercial LFAs. (A) Calibration curves (n = 3) for the Romer and 3M peanut LFA assays. (B) Calibration curves (n = 3) for the 3M gluten LFA assay.

Table 1. Analytical Parameters of the Calibration Curves Shown in Figure 3.

phone	assay	RSD (%)	R2	LOD (ppm)	IC50 (ppm)	linear range (ppm)
Android	Romer, peanut	7.0 ± 5.5	0.97	4.2	14.2	6.6–32.0
	3M, peanut	1.8 ± 1.0	0.99	6.8	13.8	9.2–20.5
	3M, Gluten	7.1 ± 3.0	0.97	688.6	1165.8	828.5–1693.1
iPhone	Romer, peanut	2.3 ± 2.0	0.94	3.5	19.4	6.2–55.7
	3M, peanut	1.3 ± 0.6	0.97	3.4	13.6	6.1–29.9
	3M, Gluten	3.4 ± 1.5	0.97	295.6	664.9	355.5–1680.7

Optimization of Peptide Quantification from LFAs

Peanut extracts from the sample pads, wick pads, and test lines of the LFAs were analyzed by LC–MS to determine which section of the LFA is best suited for LC–MS analyses. The top ten best performing peptides were selected for further assay optimization (Figure 4). For this selection, a cut-off was used whereby the peak area of the wick pad and/or the sample pad needed to be at least 1 × 10⁶ counts^–s. Test-line extracts did not produce significant signal for LC–MS analyses—likely due to the limited amount of sample present on the test line. The peak area obtained for sample pad extracts was considerably higher than the peak areas obtained for the wick pad for all selected peptides (Figure 4A). The peak area CVs for the sample pads were also considerably lower as the wick pad CVs for nine of the 10 peptides tested with various CVs well below 10% (Figure 4B). Thus, sample pads were used for the remainder of extractions.

Figure 4.

Optimization of the LC–MS extraction protocol. (A) Average peak area values (n = 3) and (B) average CVs for sample and wick-pad extracts of the top ten identified peanut peptides.

LC–MS Assay Performance

LC–MS performance for the quantification of peanut and gluten protein directly from the sample pads of the three commercial LFA assays was tested using the best performing peptides for each assay (TANDLNLLILR for peanut and LEGSDALSTR for gluten determination). An excellent linear range was obtained for all three calibration curves ranging from 2000 to ∼8 ppm for both peanut LFA types and 4000 to ∼2 ppm for the gluten LFAs with R² > 0.95 for the peanut and R² > 0.93 for the gluten assays (Figure 5A–C). The LOQs for all assays were quite close to one another (7–8 ppm). Notably for the peanut LFAs, the LC–MS LOQs were slightly above the LOD of the smartphone, while for gluten, the LC–MS LOQ was over 35-fold below the smartphone LOD. All three LC–MS assays showed good analytical performance when compared with the performance of smartphone-based analyses using raw file analyses (Table 1; Figure 3). Particularly, LC–MS assays had LODs well below the LODs determined for the smartphone assays as well as the visual LODs reported by the manufacturer.

Figure 5.

Direct quantification of allergenic peptides from LFAs by LC–MS. (A) Calibration curve (n = 3) constructed from sample pad extracts from the Romer peanut assay. (B) Calibration curve (n = 3) constructed from sample pad extracts from the 3M peanut assay. (C) Calibration curve (n = 3) constructed from sample pad extracts from the 3M gluten assay. Inset figures show the peaks of the quantitative and qualitative transitions at the LOD (S/N > 20 for all graphs). (D) RSDs (n ≥ 18) calculated for the calibration curves shown in (A–C).

Regarding repeatability, the LC–MS quantification of the Romer peanut LFAs performed well (CV = 8.5 ± 6.3%) shortly followed by the 3M assay (CV = 15.2 ± 4.4%; Figure 5D). The gluten assay performance was slightly lower in terms of repeatability (CV = 28 ± 8.0%). As such, the linear range, LOD, LOQ, and RSD parameters of the LC–MS assays for peanut quantification directly from LFAs all comply with the AOAC Standard Method Performance Requirements (SMPRs) for Detection and Quantitation of Selected Food Allergens (AOAC SMPR 2016.002). For gluten determination, the LOQ and CV are slightly above the recommended limit in the AOAC SMPR for Quantitation of Wheat, Rye, and Barley and would require further optimization. However, the LC–MS assays are clearly able to detect false negatives from LFAs due to the limited sensitivity of these LFAs since clear peaks can be distinguished well below the LFA LODs and the AOAC recommended LODs (Figure 5A–C; inset).

Moreover, a protein BLAST search shows that LEGSDALSTR is a unique identifier for wheat gluten (Triticum aestivum, Triticum spelta, and Triticum dicoccoides) except for one hit for a hypothetical Escherichia coli protein (WP_161424595.1; NCBI), while TANDLNLLILR is a unique identifier for cultivated peanut and wild peanut protein (both of which carry the peanut allergens).⁸ Thus, false positives caused by antibody cross reactivity can be identified with this method. Moreover, untargeted proteomic strategies can be applied on the extract if no target peptides are detected by this targeted LC–MRM–MS assay to determine which proteins are present in the sample. This may help to identify potential candidates in the food matrix causing cross reactivity or identify the presence of unexpected proteins or absence of any protein, which may indicate that the test was performed incorrectly.

Conclusions

The results presented herein clearly show that 16-bit raw intensity values from DNG images produce superior results for smartphone-based color quantification when compared with 8-bit RGB channels and gray-scale values from JPEG images. Moreover, the method reached similar LODs as observed with the naked eye, which is an improvement to previous smartphone-based quantification.¹⁶ This is likely because the raw image format remains unprocessed, contains less artefacts, and has an improved dynamic range. Ratiometric ambient light corrections performed particularly well when applied to the raw image format, and it was demonstrated that such corrections can be successfully used to correct for ambient light variation. Direct protein extraction from LFAs and subsequent quantification by LC-MRM-MS showed great promise as a follow-up analysis, confirming not only the presence of the suspected contaminant but also verifying the credential of the screening assay itself in the process. Combining such on-site objective quantification via smartphone LFA analyses with LC–MRM–MS confirmation of LFA results can eliminate uncertainty regarding allergen content. This can greatly improve consumer trust in allergen-free products, particularly since a system can be designed whereby LFAs are used for quick on-site confirmation of contaminant absence. Furthermore, an app can be developed with a built-in system enabling programmed random requests to send the used LFA (negative or positive) for laboratory-based LC–MRM–MS analyses. Such a system can help to ensure LFA test results are genuine negatives or positives (no false negative due to LFA sensitivity issues or incorrect/fraudulent use of the LFA and no false positives due to cross reactivity issues). Obtaining LC–MRM–MS results for LFA validation would likely take a day or two but would only be done sporadically (e.g., once every 100 LFA tests) to control the rapid LFA testing procedure and ensure maximum food safety while enabling very rapid and cost-efficient testing for the additional 99 tests. Moreover, the principle is applicable to any other protein-based target detection and has excellent potential for other applications in food safety and even medical fields where it may be desirable to have a means of monitoring whether LFAs were correctly used and interpreted. This approach also constitutes a good value proposition as it can create a faster and less costly way to obtain product certification for food exporters/producers, who must often wait for days to obtain certification via costly lab-based LC–MS testing. Future research in this field will focus on combining the raw image analyses with the automated recognition and registration of LFA tests performed and automated control for the selection of LFAs to be sent for follow-up LC–MS analysis.

Supporting Information Available

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.analchem.2c03000.

• Workflow, DNG and JPG assay images, and data analyses pipeline (PDF)

Author Contributions

The manuscript was written through contributions of all authors. All authors have given approval to the final version of the manuscript.

The authors declare no competing financial interest.