A frequent goal of chemical forensic analyses is to select a panel of diagnostic chemical features─colloquially termed a chemical fingerprint─that can predict the presence of a source in a novel sample. However, most of the developed chemical fingerprinting workflows are qualitative in nature. Herein, we report on a quantitative machine learning workflow. Grab samples (n = 51) were collected from five chemical sources, including agricultural runoff, headwaters, livestock manure, (sub)urban runoff, and municipal wastewater. Support vector classification was used to select the top 10, 25, 50, and 100 chemical features that best discriminate each source from all others. The cross-validation balanced accuracy was 92-100% for all sources (n = 1,000 iterations). When screening for diagnostic features from each source in samples collected from four local creeks, presence probabilities were low for all sources, except for wastewater at two downstream locations in a single creek. Upon closer investigation, a wastewater treatment facility was located ∼3 km upstream of the nearest sample location. In addition, using simulated in silico mixtures, the workflow can distinguish presence and absence of some sources at 10,000-fold dilutions. These results strongly suggest that this workflow can select diagnostic subsets of chemical features that can be used to quantitatively predict the presence/absence of various sources at trace levels in the environment.
Keywords: chemical fingerprinting; chemical forensics; high-resolution mass spectrometry; machine learning; multivariate analysis; non-target chemical analysis.