Article pubs.acs.org/ac
Correlation Queries for Mass Spectrometry Imaging Frank Suits,*,† Thomas E. Fehniger,‡,§ Á kos Végvári,§,∥ György Marko-Varga,§,⊥ and Peter Horvatovich⊗,#,○ †
IBM T. J. Watson Research Center, P.O. Box 218, Yorktown Heights, New York 10598, United States Institute of Clinical Medicine, Tallinn University of Technology, Akadeemia tee 15, 12618 Tallinn, Estonia § Clinical Protein Science & Imaging, Biomedical Center, Department of Measurement Technology and Industrial Electrical Engineering, Lund University, BMC C13, SE-221 84 Lund, Sweden ∥ CREATE Health, Lund University, Lund, Sweden ⊥ First Department of Surgery, Tokyo Medical University, Tokyo, Japan ⊗ Department of Analytical Biochemistry, University of Groningen, Antonius Deusinglaan 1, The Netherlands # Netherlands Bioinformatics Centre, Geert Grooteplein 28, 6525 GA Nijmegen, The Netherlands ○ Netherlands Proteomics Centre, Padualaan 8, 3584 CH Utrecht, The Netherlands ‡
S Supporting Information *
ABSTRACT: Mass spectrometry imaging (MSI) generates large volumetric data sets consisting of mass to charge ratio (m/z), ion current, and x,y coordinate location. These data sets usually serve limited purposes centered on measuring the distribution of a small set of ions with known m/z. Such earmarked queries consider only a fraction of the full mass spectrum captured, and there are few tools to assist the exploration of the remaining volume of unknown data in terms of demonstrating similarity or discordance in tissue compartment distribution patterns. Here we present a novel, interactive approach to extract information from MSI data that relies on precalculated data structures to perform queries of large data sets with a typical laptop. We have devised methods to query the full volume to find new m/z values of potential interest based on similarity to biological structures or to the spatial distribution of known ions. We describe these query methods in detail and provide examples demonstrating the power of the methods to “discover” m/z values of ions that have such potentially interesting correlations. The “discovered“ ions may be further correlated with either positional locations or the coincident distribution of other ions using successive queries. Finally, we show it is possible to gain insight to the fragmentation pattern of the parent molecule from such correlations. The ability to discover new ions of interest in the unknown bulk of an MSI data set offers the potential to further our understanding of biological and physiological processes related to health and disease.
M
combination of high spatial (scanning) resolution with high mass spectral resolution can produce data files tens of gigabytes in size for a single sample scan, and there is often little a priori knowledge of the distribution or identity of specific m/z peaks. Aside from a handful of molecular masses known from experimental design, little is known about the identities or levels of abundance of the ions corresponding to the thousands of peaks in the generated data set. Furthermore, there is a paucity of analysis software that can provide evidence of the distribution and statistical associations of a given ion with other ions. MSI today is thus very good at showing distribution patterns of single ions but fails to provide needed insight into how they are related
ass spectrometry imaging (MSI) provides an effective means to measure the distribution of ions desorbed by tissue samples such as those used in standard histology analyses. The use of a scanning stage to sample at contiguous registration points makes it possible to generate maps of mass spectra at a point-by-point basis throughout the entire tissue sample area. Although the intensity of a given mass spectral peak can be influenced by many factors, the visualization of individual peaks in two-dimensional space provides an image corresponding to the peak that can relate to specific histological compartment spaces. MSI therefore has widespread usage throughout the community with academic, clinical, and commercial applications leading to a new era of method developments in detection, analysis, and statistical association.1−3 There is a common issue in MSI that an abundance of data are generated for a large number of m/z values in a single run. The © 2013 American Chemical Society
Received: December 17, 2012 Accepted: March 28, 2013 Published: March 28, 2013 4398
dx.doi.org/10.1021/ac303658t | Anal. Chem. 2013, 85, 4398−4404
Analytical Chemistry
Article
Figure 1. Flowchart showing stages in the processing of data from a MALDI-MSI sample scan to produce the three large, binary structures a, b, c at the bottom.
into compositionally similar regions. Finally, Seeley et al.12 provide a perspective on the challenges of performing MSI on human tissue and in particular what is required to discover new molecules of interest, i.e., discovery mode. This paper includes aspects of prior work, but the goals are fundamentally different. Experiments involving MSI on tissue samples often involve prior knowledge of expected biological structures within the sample, plus a list of expected analyte m/z values from, for example, known biological molecules or pharmaceutical compounds. The work reported here focuses on methods that leverage this a priori information to automatically generate results that can be directly examined by the experimentalist. The goal is for the experimentalist to input relevant knowledge such as known m/z values of interest and known biological structures within the tissue sample and to use this knowledge as an abstract query on the full MSI data set that returns, as an answer, a ranked list of m/z values and corresponding images of the tissue that spatially correlate with the known analytes, the known structures, or some combination of both. This ranked list of images can easily be scanned by the experimentalist to determine if any are interesting enough to warrant follow up study, and since each image represents a unique and previously unknown m/z value, it is operating in discovery mode whereby the entire volume of MSI data is automatically mined for new analytes of interest and directly
to all the other ions detected in the sample. Although the work in this paper is based on MSI using matrix-assisted laser desorption and ionization (MALDI-MSI), the methodology would apply to other MSI techniques as well. MSI studies often target a few known ions and effectively ignore the remainder of the entire global data set of m/z peaks, though there has been recent effort to find statistical associations within the data sets in a form of discovery mode. McCombie et al.4 did early work to find spatially correlated behavior among analytes in MALDI-MSI data. Their method works either with preselected regions of interest, in which principal components analysis (PCA) identifies patterns of m/z values that are common across the pixels, or it can also identify spatial regions that have a consistent analyte pattern within them. Others have used hierarchical clustering5 and a statistical assessment of correlation significance6 in analyzing MSI data. Later, McDonnell et al.7 addressed the importance of data size reduction in analyzing the large volume of data produced by MALDI MSI through userdefined thresholds on m/z peaks Recently Bruand et al.8 developed a pipeline for MALDI MSI analysis that uses selected regions of interest (ROI) and statistically identifies peaks that show differential intensity within those regions. Alexandrov et al.9,10 and Trede et al.11 developed a spatially aware approach that combines the imaging aspect of MSI with the molecular detail of the mass spectra to automate the segmentation of a tissue sample 4399
dx.doi.org/10.1021/ac303658t | Anal. Chem. 2013, 85, 4398−4404
Analytical Chemistry
Article
nonzero intensity value in every scan. This 12-gigabyte binary file is easily accessed to a specific m/z range using a separate index file generated during the conversion and allows the interactive creation of slices of arbitrary thickness. Finally, we create another lossless file comprising the full set of mass spectra captured at each location in the scanned region, which allows a direct probe of the raw spectrum at each scan location. For the resolution of our Orbitrap mass-spectrometer, the 0.1 Da thickness of the slices represents a good trade off between the separation of peaks from distinct analytes and the undesirable splitting of a single peak into two adjacent slices. In most cases the ion current from only a single peak is captured in each slice, but if for a given sample there is more or less crowding of the peaks, then the thickness of the slices can be set accordingly or even allowed to vary with m/z in accordance with the resolution behavior of the mass spectrometer.15 However, for this study, we simply bin the volume of data based on the integrated ion current in small steps of m/z in order to capture an approximate representation of the range of analytes present in the sample, most of which are assumed unknown based on limited prior knowledge. We can then perform a number of different types of queries on the full data set based on combinations of input options. The first query type uses, as input, a single m/z value of a known ion and seeks other m/z values that show a similar global distribution pattern across the image, regardless of its relative intensity. Although the published methods described above for querying and segmenting MSI data sets can involve stages of normalization, peak picking, and estimates of statistical significance, our approach is much simpler and we simply perform a Pearson correlation of the query slice with all other slices in the data volume (pseudocode in the Supporting Information, page S-2). This is not only computationally efficient, but it has a key advantage that it does not require normalization, scaling, or background subtraction in an absolute sense, since each slice is scaled to its own min/max range and directly correlated with the query to find similar distributions of ion signal across the sample. The query slice could be a strong signal from a prominent ion, and it could correlate most strongly with a very weak peptide signal near the level of noise as long as its signal is clear enough to display a matching distribution. Since this method is just generating “leads” on ions of interest in discovery mode, there is no need for a formal normalization of the images to correct for complications such as ion suppression. The aim is to discover any ions that are somehow spatially correlated with a query, regardless of the absolute intensity of the signal. Similarly to the work of Bruand,8 we also allow a query mode in which the user specifies a mask indicating a known biological structure in the specimen. This mask can itself be a query, seeking m/z values that match its shape and with no corresponding input m/z value, or it can supplement an m/z query by finding other m/z values that match its distribution but only within the region specified by the mask. This lets the experimentalist apply domain knowledge at the outset of analysis, by manually constructing a mask from the histology image and merging it with known m/z values to generate a list of new m/z values that are correlated with known analytes in those regions. In summary we provide three query modes: m/z alone, mask alone, and m/z combined with mask. Note that a mask could either be a human-drawn outline of known features in the histology image, created as a registered overlay of the image, or it could be an actual histology image based on a particular dye or fluorescence measurement.
presented in a spatial form that can be interpreted in its comparison to histology images and the distribution of other analytes.
■
EXPERIMENTAL SECTION Sample Collection and Preparation. Animals. The animal tissue used in this study was obtained under ethical approval (Lund/Malmo, Sweden Ethical Committee on Animal Experiments (no. M84-05). Male rats (Wistar, Taconic, Denmark) were anesthetized and sacrificed and the lungs removed and processsed for cryostat sectioning as previously described.13 Lung sections mounted on glass slides were coated with α-cyano4-hydroxycinnamic acid (7.5 mg/mL) dissolved in 50% acetonitrile and 0.1% trifluoroacetic acid14 as the MALDI matrix. Following MALDI-MSI analysis, MALDI slides were stained using conventional hematoxylin and eosin staining and then digital images of the stained sections were acquired (ZeissMiromax scanner) at 67 000 × 81 000 pixels resolution. MALDI-MS Imaging. For the imaging mass spectrometry scans, we used a MALDI LTQ Orbitrap XL mass spectrometer (Thermo Scientific, Bremen, Germany), equipped with a 60 Hz 337 nm nitrogen pulse laser (LTB Lasertechnik Berlin GmbH, Berlin, Germany), in positive mode, sampling the tissue sections with 30 μm raster arrays without laser movement within each measuring position.14 We measured the ablated regions to be approximately 25 μm in diameter, allowing negligible overlap of samples taken at 30 μm intervals. The Fourier transform (Orbitrap) mass analyzer operated at 60 000 resolution (at 400 Da) collecting spectral data in the range of 350−2000 Da in profile mode generated by 20 laser shots at 10 μJ with automatic gain control switched off. Data Reduction and Types of Queries. In order to avoid any loss of information from the mass spectra, we work strictly with the raw, profile mode output from the mass spectrometer rather than centroid mode, in which peaks are identified in each scan and output as a single m/z, intensity pair rather than a full profile of the peak, with intensity as a function of m/z. Although the size of files is much greater in profile mode, there is no loss of information and one avoids possible rejection of faint signal peaks.15 Figure 1 shows the full processing pipeline for a scanned tissue sample. A single MSI data set from our tissue samples consists of a 408 × 498 array of MALDI scans with 30 μm spacing (a total of 230 184 sample points). We convert the raw mass spectrometry data to mzXML format with an in-house Visual Basic script using Thermo Xcalibur software (version 2.1 Build 0.40, Thermo Fisher Scientific, Waltham, MA). Another script extracts the x, y coordinates from each spectrum header to reconstruct the scan. We then convert the raw mzXML data into three large, binary structures for subsequent analysis. First, for efficient correlation calculations of a given analyte distribution with other m/z values, we create a binned version of the volume of data representing the total ion current found in each m/z slice of user-specified thickness, or 0.1 Da for examples presented in this work. One may view this representation as a stack of slices of the sample volume at each step of 0.1 Da through the range of m/z of interest, which for this study is 350− 1000 Da, or 6500 distinct slices. Although it is not practical for an experimentalist to go through each slice manually for inspection, the precalculated slices serve as the basis for the discovery-mode queries of the data volume described below. Second is a lossless binary data file containing the full set of approximately 1 billion triplets (m/z, pixel ID, intensity), corresponding to every 4400
dx.doi.org/10.1021/ac303658t | Anal. Chem. 2013, 85, 4398−4404
Analytical Chemistry
Article
Figure 2. Query by m/z value using an input m/z of 676.3 Da, with the H&E stained histology image (a), the selected m/z slice (b), and the primary correlated m/z slices on the right (c−e) with each slice labeled by its m/z value.
Figure 3. Query using a mask alone (a, left) finds correlated m/z slices (a, upper) and anticorrelated (a, lower) with decreasing correlation or anticorrelation from left to right. Each slice labeled by its m/z value. A full plot of the correlation values for all slices is shown in part b. The slices of likely interest for experimental follow up are the ones at the extreme left and right of the ordered list, allowing directed investigation of a small subset of the full list of slices.
It is important to emphasize that the response to these correlation queries is always the entire list of slices but sorted
from most correlated, to most anticorrelated; and the most interesting values are those at the extreme values of correlation. 4401
dx.doi.org/10.1021/ac303658t | Anal. Chem. 2013, 85, 4398−4404
Analytical Chemistry
Article
Figure 4. Schematic representation of a query that combines a mask with a target m/z value (676.3 Da) to find structures correlated only under the mask shown while ignoring the rest of the sample. Note that this reveals a different set of structures from the query by m/z alone (Figure 2) and mask alone (Figure 3), including m/z values that show a faint but clearly correlated appearance under the mask.
Figure 5. Close up of H&E stained histology section and corresponding slice at m/z = 650.2 Da. The green circle denotes a 3 × 3 MALDI scan location (90 μm × 90 μm) selected for the spectrum shown in Figure 6. Note the fine detail in the scanned image matches well with the histology image.
29th in the list of slices sorted by TIC and the first to show interesting structure that might serve as a fruitful query. It was thus “discovered” by a simple manual scan of the intensity-sorted list, which only took a few minutes of manual inspection of the images in order. We used this m/z value by itself as a query to find other m/z values that show a similar distribution of ion current across the entire sample. Figure 2c−f are the first four “answers” to the query in order of decreasing correlation. We ignored trivially related m/z values such as isotopes 1 Da apart, and one can see a number of similar distributions either due to other fragments from the same parent ion species or a different species with a similar distribution in the specimen and possibly a related biological function. Although this query involves an unknown initial m/z value, the same query would work well in a specimen with a known landmark ion mass such as a pharmaceutical small compound. Note also that although the input query at 676.3 Da represented a strong signal with high ion current, the correlated m/z values could have very weak ion current as long as the overall signal roughly matches the query distribution. The second type of query is based on correlations with a mask designed to match known biological structures of interest. Figure 3a shows the distributions of ions discovered to be correlated or anticorrelated with a mask plane that was drawn upon large blood vessels. The result of the query is the full ranking of likeness (most correlated) and dissimilarity (anticorrelated) of all slices. For context, Figure 3b shows the full set of correlations with the mask for all slices, sorted from most correlated on the left to most anticorrelated on the right. The m/z values worthy of potential experimental follow up are those at the extreme ends of the correlation spectrum. As with many of the queries described in this paper, there is no attempt to provide a “cut-off” value for which the correlation has no potential significance. Instead, this
Rather than apply heuristics in combination with a threshold to reduce the size of the query response, we simply provide a ranked list of all slices and allow the experimentalist to study them in order and assess their importance based on domain knowledge. During this assessment phase, the user can examine the raw spectrum of ion currents in each m/z slice to determine if it represents signals from multiple peaks or if a single peak is only partially present in the given slice. In either case, a new slice can be captured on-demand representing a custom range of m/z values that match the exact width of a selected peak using the m/ z-sorted binary file shown in Figure 1c. Again, we chose 0.1 Da as a good thickness to match our experimental conditions, but other experiments would use a larger or smaller slice thickness depending on the mass-spectrometer resolution and the density of peaks in the spectra.
■
RESULTS AND DISCUSSION In this model study we demonstrate the potential of novel, interactive queries to investigate the distribution of ions detected by MALDI-MSI analysis of an adult rat lung. A goal of our system is to make use of any existing knowledge of histology structures and m/z values to discover new features in a large MSI data set, and in this work we rely on a previous study in this lung model where we acquired experience mapping several known and unknown ion species of this tissue compartment13 One of the simplest initial queries is to sort the slices by total ion current to examine those with the strongest signal. This is a manual process of inspecting those with the strongest signal, but it benefits from the preslicing of the data and signal ranking, making it effectively a query by intensity. For example, an ion mass value at m/z = 676.3 Da shows a strong signal based on total ion current (TIC) within the lung tissue sample (Figure 2a,b). This m/z value is the 4402
dx.doi.org/10.1021/ac303658t | Anal. Chem. 2013, 85, 4398−4404
Analytical Chemistry
Article
correlated with the query m/z, and one can see that they constitute only a small subset of peaks in that range, including peaks that have much higher ion current and are often very close to one of the correlated peaks. This demonstrates that the resolution of 0.1 Da in this case was fine enough to extract faint signals based on correlation and provide adequate separation from neighboring peaks that do not have a similar spatial distribution while remaining computationally efficient. The ions discovered by query can also provide insight into the parent species based on the pattern of m/z values that show correlation in a query. Figure 7 shows 22 of the m/z values that
simple mask-based query yields a ranked list of the slices of interest, which the researcher can manually inspect either by studying the distributions or generating additional queries based on the new m/z values. The queries by m/z in Figure 2 and by mask in Figure 3 can be combined into a joint query as shown in Figure 4, where the distribution of m/z = 676.3 Da is matched with other slices only under the region of the mask. This has the important benefit of being blind to strong ion current outside the mask that would result in a weak correlation by m/z alone. As a result, slices with very small ion current in the region of the mask may still correlate well if the distribution of the queried ion signal matches that of the target m/z. The queries described above benefit from the sliced representation of the MSI data set for efficient and interactive correlation calculations, but additional insights can be obtained through the exploration of the data sets in a series of related queries and refinements to isolate a peak of interest, where each different query relies on precomputed data structures for good performance. Once an interesting m/z value has been discovered by the query process, the slice can be further investigated by refining the slice m/z range to isolate a single peak in case there is more than one peak in the spectrum contained by the slice. This quick reslicing of the full data set is made computationally efficient by the large data structure representing the sorted triplet of m/z, ion-current, scan location values, since the slice can be directly constructed from a contiguous subset of values in the file ordered by m/z. When the resulting m/z image is compared to the histology slice, selected mass spectra can be chosen from a single probe point or, to improve detection of fainter signal, a 3 × 3 pixel region in the image, as shown in Figure 5. The ability to extract spectra quickly based on scan location and aggregate them to improve the signal-to-noise ratio is due to the third data structure shown in Figure 1c, consisting of the binary m/z scans for each scan location stored in a single, indexed file for quick access. Figure 6 shows a section of the full m/z spectrum probed at the circled region of Figure 5 to focus on m/z values that were “discovered” in a query using m/z = 650.2 Da. Each blue triangle marker along the x-axis represents an m/z value that was spatially
Figure 7. Schematic plot of “discovered” ions and their steps in m/z, which provides insight into the fragmentation pattern and molecular composition of the corresponding peaks. All of the smaller steps are 1 Da, and any m/z shifts different from 1 are labeled on the plot. The circle indicates the query m/z (650.2 Da).
match a query with target m/z = 650.2 Da, arranged in order of m/z so the steps between them are evident. Many of the steps are 1 Da, presumably corresponding to isotopes, but other steps have values of 10, 13, and 24 Da, suggestive of a fragmentation pattern. The final step of 29.8 Da may be due to an unrelated ion. Note that these are not simply a set of m/z values found at a given location: they are a set that appear codistributed in the specimen with the target m/z value of 650.2 Da. They may all be different fragments of a parent species, or they may be biologically linked analytes, or they may simply be unrelated species that happen to be spatially correlated. Additional investigation is needed to identify the m/z values discovered by the queries described in this work, but the interactive exploration and discovery of such potentially interesting m/z values is made possible by the query methods and data structures described here.
■
CONCLUSIONS We have described methods for querying large imaging mass spectrometry data sets to discover m/z values of potential biological interest in an interactive and automated manner. A key to the practicality of this approach is the precalculation of large data structures that allow typical personal computers to perform the queries interactively. This work aims to find possibly useful information in a discovery-mode exploration of an existing data set, to gain insights that would otherwise involve an impractically slow manual effort. We do not attempt to characterize the statistical significance of the discovered m/z values but merely offer them as leads that might warrant a more focused and
Figure 6. Plot of the spectrum through 3 × 3 probe location shown in Figure 5, with triangle markers indicating m/z values that show correlation with the original query m/z = 650.2 Da. Although there is a dense forest of nearby peaks at that location, only those marked with a triangle appeared as well-correlated with 650.2 Da. 4403
dx.doi.org/10.1021/ac303658t | Anal. Chem. 2013, 85, 4398−4404
Analytical Chemistry
Article
quantitative investigation into their biological importance in the experiment. Regardless, they provide a means to extract more information from large, mass spectrometry imaging data sets than might otherwise be known based on only a handful of a priori m/z values of interest.
■
ASSOCIATED CONTENT
S Supporting Information *
Additional information as noted in text. This material is available free of charge via the Internet at http://pubs.acs.org.
■
AUTHOR INFORMATION
Corresponding Author
*Phone: +01-914-945-1294. E-mail:
[email protected]. Notes
The authors declare no competing financial interest.
■
ACKNOWLEDGMENTS The Groningen laboratory gratefully acknowledges the following grants: BioRange (NBIC) 2.2.3; NPC (Netherlands Proteomics Centre) Bsik03015. The authors at Lund University are grateful for funding support from the Inga-Britt & Arne Lundberg Foundation, the Knut and Alice Wallenberg Foundation, the Crafoord Foundation, and the Vinnova Support 2011-03926 for CREATE Health.
■
REFERENCES
(1) Ait-Belkacem, R.; Sellami, L.; Villard, C.; DePauw, E.; Calligaris, D.; Lafitte, D. Trends Biotechnol. 2012, 30, 466−474. (2) Fonville, J. M.; Carter, C.; Cloarec, O.; Nicholson, J. K.; Lindon, J. C.; Bunch, J.; Holmes, E. Anal. Chem. 2012, 84, 1310−1319. (3) Zimmerman, T. A.; Debois, D.; Mazzucchelli, G.; Bertrand, V.; De Pauw-Gillet, M. C.; De Pauw, E. Anal. Chem. 2011, 83, 6090−6097. (4) McCombie, G.; Staab, D.; Stoeckli, M.; Knochenmuss, R. Anal. Chem. 2005, 77, 6118−6124. (5) Deininger, S. O.; Ebert, M. P.; Futterer, A.; Gerhard, M.; Rocken, C. J. Proteome Res. 2008, 7, 5230−5236. (6) McDonnell, L. A.; van Remoortere, A.; van Zeijl, R. J. M.; Deelder, A. M. J. Proteome Res. 2008, 7, 3619−3627. (7) McDonnell, L. A.; van Remoortere, A.; de Velde, N.; van Zeijl, R. J.; Deelder, A. M. J. Am. Soc. Mass Spectrom. 2010, 21, 1969−1978. (8) Bruand, J.; Sistla, S.; Meriaux, C.; Dorrestein, P. C.; Gaasterland, T.; Ghassemian, M.; Wisztorski, M.; Fournier, I.; Salzet, M.; Macagno, E.; Bafna, V. J. Proteome Res. 2011, 10, 1915−1928. (9) Alexandrov, T.; Becker, M.; Deininger, S.-O.; Ernst, G. n.; Wehder, L.; Grasmair, M.; von Eggeling, F.; Thiele, H.; Maass, P. J. Proteome Res. 2010, 9, 6535−6546. (10) Alexandrov, T.; Kobarg, J. H. Bioinformatics 2011, 27, i230−238. (11) Trede, D.; Schiffler, S.; Becker, M.; Wirtz, S.; Steinhorst, K.; Strehlow, J.; Aichler, M.; Kobarg, J. H.; Oetjen, J.; Dyatlov, A.; Heldmann, S.; Walch, A.; Thiele, H.; Maass, P.; Alexandrov, T. Anal. Chem. 2012, 84, 6079−6087. (12) Seeley, E. H.; Caprioli, R. M. Trends Biotechnol. 2011, 29, 136− 143. (13) Nilsson, A.; Fehniger, T. E.; Gustavsson, L.; Andersson, M.; Kenne, K.; Marko-Varga, G.; Andrén, P. E. PLoS One 2010, 5 (7), e11411. (14) Vegvari, A.; Fehniger, T. E.; Gustavsson, L.; Nilsson, A.; Andren, P. E.; Kenne, K.; Nilsson, J.; Laurell, T.; Marko-Varga, G. J. Proteomics 2010, 73, 1270−1278. (15) Suits, F.; Hoekman, B.; Rosenling, T.; Bischoff, R.; Horvatovich, P. Anal. Chem. 2011, 83, 7786−7794.
4404
dx.doi.org/10.1021/ac303658t | Anal. Chem. 2013, 85, 4398−4404
