Emerging investigator series: automated single-nanoparticle quantification and classification: a holistic study of particles into and out of wastewater treatment plants in Switzerland

Single particle inductively coupled plasma time-of-flight mass spectrometry (sp-ICP-TOFMS), in combination with online microdroplet calibration, allows for the determination of particle number concentrations (PNCs) and the amount (i.e. mass) of ICP-MS-accessible elements in individual particles. Because sp-ICP-TOFMS analyses of environmental samples produce rich datasets composed of both single-metal nanoparticles (smNPs) and many types of multi-metal NPs (mmNPs), interpretation of these data is well suited to automated analysis schemes. Here, we present a new data analysis approach that includes: 1. automatic particle detection and elemental mass determinations based on online microdroplet calibration, 2. correction of false (randomly occurring) multi-metal associations caused by measurement of coincident but distinct NPs, and 3. unsupervised clustering analysis of mmNPs to identify unique classes of NPs based on their element compositions. To demonstrate the potential of our approach, we analyzed water samples collected from the influent and effluent of five wastewater treatment plants (WWTPs) across Switzerland. We determined elemental masses in individual NPs, as well as PNCs, to estimate the NP removal efficiencies of the individual WWTPs. From WWTP samples collected at two points in time, we found an average of 90% and 94% removal efficiencies of single-metal and multi-metal NPs, respectively. Between 5% to 27% of detected NPs were multi-metal; the most abundant particle types were those rich in Ce–La, Fe–Al, Ti–Zr, and Zn–Cu. Through hierarchical clustering, we identified NP classes conserved across all WWTPs, as well as particle types that are unique to one or a few WWTPs. These uniquely occurring particle types may represent point sources of anthropogenic NPs. We describe the utility of clustering analysis of mmNPs for identifying natural, geogenic NPs, and also for the discovery of new, potentially anthropogenic, NP targets.


Table of Contents
Operating parameters Table S2 List of selected elements and the isotopes used for sp-ICP-TOFMS analysis Table S3 Figure S1 Single-particle critical value expression parameters WWTP Sampling sites Figure S2 Response curve of the instrument Figure S3 Example of hpCC for Ag-Sn mmNPs Figure S4 Graphical illustration of the two stage hierarchical clustering Figure  Insight into the Ce-La cluster (May 2019) Figure S11 Comparing the La to Ce mass ratio of the particle events

Semi-quantitative mass analysis
As shown in equation S1, the absolute sensitivity of each element ( ) is calculated, where is the droplet signal, is the abundance of used isotope(s), is the mass of element of interest in the droplet and is the molecular weight of the element in the droplet. This formula used to calculate for the absolute sensitivity of the element in the droplet and further predict the ones missing as shown in Fig S1. (S1)  in the droplet solutions are in red. The green line shows the linear regression of the known response in each low, median, and high mass region.

Hetero-particle coincidence correction (hpCC)
The first step in our hpCC algorithm is to group every mmNP that have identical element fingerprints and sort the fingerprints from those with the most to those with the fewest number of elements. Starting with the fingerprint with the most elements, we define this particle fingerprint ( ) as its representative set of elements, as shown in equation S2, in which are the fingerprint elements. Next, from our pool 1 , 2 ,… of found smNP and mmNP types we find possible combinations of two precursor particles ( ) that 1 , 2 together would produce a coincident particle with all fingerprint elements in (see equation S3), and in which no fingerprint elements are conserved between and (see equation S4). Once these and data points, and is predicted number of coincident and events.
( 1 , 2 ) 1 2 If the is greater than 1, then we attempt to identify which of the measured particles with element ( 1 , 2 ) signature are likely due to particle coincidence (i.e. and occurring together) rather than due to 1 2 true multi-metal composition. To accomplish this, we define a hypothetical selection criterion and a scoring method based on the ratios of element signals in the measured mmNPs with composition . Let us consider the set of individual particle signals that all have the identical fingerprint, , to be For scoring individual particle signals, we first find the element with the largest = { 1 , 2 ,… } .
median signal (E max ) in the set of particle signals, , as shown in equation S6. To normalize the element signals within each particle, we then divide the element signals of all elements in each particle by the signal of the E max in the particle, i.e. . A sole normalized value for each particle in the set is defined as and is the sum of the normalized element signals in each particle, as shown in equation S7. A score of how similar each particle is to the overall group ( ) is defined as the distance of the particle's from the median values of the set, , as shown in equation S8. Particles with the largest values are least likely to be part of the true mmNP set and are therefore identified as particle coincident events. A total of particles with the highest scores are then removed from the ( 1 , 2 ) set and broken down into their composite and particles. In Fig S3, we provide an example of mmNP 1 2 similarity scoring and the subsequent breakdown of Ag/Sn mmNP signals into coincident and noncoincident events. Once all possible coincident particle combinations and their coincidence probabilities are calculated for particles with a given set of fingerprint elements, i.e. , the hpCC process is repeated for mmNP signals with successively fewer numbers of fingerprint elements. hpCC is complete when all unique mmNP signals are divided into coincident particle and non-coincident particle fractions.
Example results from hpCC are shown in Fig S3. Hetero-particle coincident events occur when two or more non-identical discrete particles are present in the plasma simultaneously; the likelihood of these events occurs follows Poisson statistics. Through sample dilution one can significantly reduce the chance of coincidence: If a sample is diluted '' '' times then all nanoparticles of interest are also diluted ''d'' times; however, the probability of a coincident event reduces as the square of the dilution factor, i.e. as '' '', 2 see equation S11. Because coincident NP events are reduced more rapidly than individual NP events, we can use dilution experiments to check the efficacy of our hpCC algorithm to assess whether the predicted true mmNP are also present in the diluted sample. Likewise, we can assess whether the predicted coincident events are absent in the diluted sample. a)The scatter plot shows the Ag-Sn coincident events at two dilutions, where green triangles are true mmNPs found in the higher dilution (100-times dilution) sample, blue rectangles are selected as true mmNPs in the 10-times dilution and red circles are selected as coincident events of Ag and Sn single-metal particles. b) the measured and predicted number of Ag-Sn coincident events (at 100x dilution, no coincident events were predicted). C) Scores of each individual Ag-Sn mmNP; particles that had scores more than 0.54 are excluded as coincident particle events.
In Fig. S3, we plot the masses of Ag and Sn in multi-metal signals at two different dilutions. At a dilution factor of just 10-times, we predict ~5 % of the multi-metal signals are the result of coincident particle events. Based on equations 6-8 for selection criteria, we identify a subset of the measured multimetal NPs that are likely to be coincident particle events and then exclude these signals from the pool of true multi-metal Ag-Sn NPs. With 100-times dilution, we predict no coincident particle events between Ag and Sn: all Ag-Sn events present in the 100-times dilution sample are likely true mmNPs. Through comparison of scatter plot of the masses of Ag vs Sn in individual measured particles, we demonstrate that our selection criteria used to decide which multi-metal signals are "true" and which are "false" is performing as expected. Our analysis of measured Ag-Sn indicates that high Ag:Sn ratio signals are likely false and this is also found in the 100-time dilution samples. No true mmNPs with Ag mass above 1 fg are found, just as predicted. Importantly, hpCC enables improved measurement statistics for mmNPs because it allows for higher number of events to be counted per unit time. Fig. S4. Graphical illustration of the two-stage hierarchical clustering (HC) of mmNPs. In the first step, HC is performed on individual samples. Then, a representative for each cluster from each sample was taken as described in the manuscript. These representative mmNP types are clustered again to find inter-sample clusters and unique mmNP types.   S6. Average equivalent spherical diameter detection limits of oxide forms of elements measured for the 10 analyses of the different WWTP samples (I1-I5 and E1-E5) from the Nov. sampling. For platinumgroup and noble metals, no oxide forms are reported. The detection limits, which are equivalent to the critical values used to identify single-particle signals, are a function of background concentrations, element sensitivities that depend on the analyte isotope, and densities of the given oxides and metals. For example, the elevated detection limit of BaO compared to La 2 O 3 comes from the fact that Ba was quantified using a minor isotope ( 137 Ba) and the barium background level is, in general, more than 10 times higher than that of lanthanum in the WWTP samples. More details on all the element sensitivities, background count rates, and critical values are provided in the SI excel spreadsheet.