Improving the identification of the source of faecal pollution in water using a modelling approach: From multi-source to aged and diluted samples

The last decades have seen the development of several source tracking (ST) markers to determine the source of pollution in water, but none of them show 100% specificity and sensitivity. Thus, a combination of several markers might provide a more accurate classification. In this study Ichnaea® software was improved to generate predictive models, taking into account ST marker decay rates and dilution factors to reflect the complexity of ecosystems. A total of 106 samples from 4 sources were collected in 5 European regions and 30 faecal indicators and ST markers were evaluated, including E. coli, enterococci, clostridia, bifidobacteria, somatic coliphages, host-specific bacteria, human viruses, host mitochondrial DNA, host-specific bacteriophages and artificial sweeteners. Models based on linear discriminant analysis (LDA) able to distinguish between human and non-human faecal pollution and identify faecal pollution of several origins were developed and tested with 36 additional laboratory-made samples. Almost all the ST markers showed the potential to correctly target their host in the 5 areas, although some were equivalent and redundant. The LDA-based models developed with fresh faecal samples were able to differentiate between human and non-human pollution with 98.1% accuracy in leave-one-out cross-validation (LOOCV) when using 2 molecular human ST markers (HF183 and HMBif), whereas 3 variables resulted in 100% correct classification. With 5 variables the model correctly classified all the fresh faecal samples from 4 different sources. Ichnaea® is a machine-learning software developed to improve the classification of the faecal pollution source in water, including in complex samples. In this project the models were developed using samples from a broad geographical area, but they can be tailored to determine the source of faecal pollution for any user. [Display omitted] •Samples from 5 geographical sources were analysed with 30 faecal markers and indicators.•A machine learning software was used to develop faecal source discriminant models.•An in-silico matrix was generated using faecal samples, adding dilution and inactivation.•LDA models’ output was a combination of markers able to improve the accuracy of classification.•Models using between 2 and 5 source tracking markers can achieve LOOCV accuracies of over 95%.