Detection and identification of microplastics in biota using Nile red and machine learning : validation of an innovative, cost-effective approach

1. Introduction Pollution of the oceans by plastics has been recognised as a serious global environmental problem. Microplastics (MPs) are of special concern as it has been shown these are ingested by a wide range of marine species from different trophic levels [1]. Despite the ubiquity of plastics in the oceans, there is only limited knowledge about temporal and spatial trends in the distribution and abundance of these plastics. Little is also known about environmental implications of MPs ingestion, coupled to the toxicological relevance of different MP polymers and their chemical composition [1]. Unanswered questions concerning MP abundance, composition, distribution and fate in the marine environment emphasize the need for standardised, reliable, cost- and time-effective monitoring procedures. This will allow to accurately and routinely determine the abundance, composition and distribution of MPs, so that effective management strategies can be implemented. An innovative approach for MP analysis in marine biota was developed, thereby combining the advantages of both high-throughput screening and automation. We evaluated the method in mussel samples and different fish gastrointestinal tracts (GIT) samples. The presented method combines RGB-color quantification of emitted fluorescence of Nile red (NR)-tagged MPs with a simplified supervised machine learning (ML) classification model, which has been developed in an open source environment [2]. Our approach further enhanced the Nile red technique for the fast detection of MPs and included the identification of the polymer types. 2. Materials and methods For MP analysis in biota, a two-step digestion was applied within the sample preparation to remove all lipid-rich animal tissue and plant and algae materials present. An alkaline digestion with potassium hydroxide (KOH, 10%) and an oxidative digestion with hydrogen peroxide (H202, 15%) was implemented, each for 48h at 50°C on a magnetic stirring plate. In between both digestion steps, The sample was filtered over a stainless steel filter and sonicated to loosen all particles adhering to the filter surface. For sediment containing samples such as fish GITs, two consecutive density separation steps of 24h each were performed, using a saturated NaI solution. The sample was then filtered over a PTFE-filter, stained with the fluorescent dye Nile red (1µg/ml acetone), and left to dry for 24h. The automated detection of MPs and the polymer type identification was based on two machine learning decision models (Figure 1), which used recursive binary splits through simple decision rules inferred from emission spectra features [2]. Both models were trained and validated using stained reference MPs. The first model predicted with high accuracy whether a particle is plastic or of natural origin. The second model allowed to identify the polymer type of MPs. Both models use RGB colour data, extracted from the stained particles photographed through a fluorescence microscope under blue, green and UV filters (Figure 1). The accuracy and precision of the analytical method was calculated using recovery tests [3]. Six MP types were spiked into the mussels and fish GITs, varying in the combination of polymer type (polyacrylonitrile -PAN, polyethylene terephthalate - PET, polypropylene - PP, polystyrene - PS and polyvinyl chloride - PVC), size (250 – 1000 µm), and shape (particle/fibre) (Figure 2). Next to accuracy and precision, the validation included the determination of the limit of quantification (LOQ), selectivity, specificity, and robustness. Based on procedural blanks, the LOQ was determined below which no quantification of particles in a sample could be performed. Selectivity and specificity were assessed based on the performance of both ML models to detect plastic particles and to identify the polymer types, respectively. To assess the robustness of the overall procedure, sample type was varied across the experiments. 3. Results and discussion Figure. 2 shows MPs spiked in a mussel sample, analysed under different fluorescence filters. MP detection and identification is done through the ML models that use RGB-color quantification to classify the particles, but the MPs can also be visually distinguished based on colouration and fluorescence intensity. As can be deduced, the combination of the emitted fluorescence of particles under all three filters is unique for each particle type assessed in our analyis, not just in terms of fluorescence intensity but also in terms of colouration. This approach solved the issue of co-staining of natural materials by Nile Red and allowed for the accurate, precise and selective determination and identification of MPs in diverse biota matrices. 4. Conclusions This unique approach of high-throughput screening and machine learning automation, proved to be promising for the cost- and time-effective routine analysis of MPs in mussel and fish GITs in a simple, yet reliable way. 5. References [1] Bergmann, M., Gutow, L., & Klages, M. (2015). Marine anthropogenic litter. Heidelberg, Germany: Springer Nature. 447 p. [2] Meyers, N., Catarino, A. I., Declerq, A. M., Brenan, A., Devriese, L., Vandegehuchte, M., De Witte, B., Janssen, C., Everaert, G. 2021: Microplastic detection and identification by Nile red staining: towards a semi-automated, cost- and time-effective technique. Manuscript submitted for publication. [3] Claessens, M., De Meester, S., Van Landuyt, L., De Clerck, K., & Janssen, C. 2011. Occurrence and distribution of microplastics in marine sediments along the Belgian coast. Marine pollution bulletin, 62(10), 2199-2204. Acknowledgement – This work was done within the JPI Oceans project Andromeda and funded by BELSPO. We want to thank the lab technicians of the Marine Analytical Laboratory of ILVO and of the Ocean & Human Health team of VLIZ. The assistance provided was greatly appreciated