Anal Bioanal Chem DOI 10.1007/s00216-014-7686-y
FEATURE ARTICLE
AMDIS in the Chemical Weapons Convention W. Gary Mallard
# Springer-Verlag Berlin Heidelberg 2014
Introduction The inspection regime under the Chemical Weapons Convention (CWC) requires both the ability to identify low levels of chemical agents, precursors, and decomposition products in complex mixtures while at the same time not exposing the inspection site to the risk of loss of confidential information related to the chemical composition of the sample. The method of choice for such detection of low levels of chemical agents is gas chromatography separation with mass spectral identification (GC/MS), but the usual method of such analysis—an extensive examination of the GC/MS data file for spectra matching the compounds of interest—is both time-consuming, prone to missing low-level components, and difficult to ensure that chemicals not relevant to the CWC are not discovered. To deal with these problems, AMDIS (Automatic Mass spectral Deconvolution and Identification System), a method of automatically finding all of the distinct chemical components in the GC/MS data file and then compare them to a library of spectra of chemicals relevant to the CWC was developed.
Fundamental algorithms of AMDIS The basic concept behind AMDIS [1] is to examine the GC/ MS data file and find all instances of ion maximization; then, for each of these cases where a maximum is found, to find all Published in the topical collection Analysis of Chemicals Relevant to the Chemical Weapons Convention with guest editors Marc-Michael Blum and R. V. S. Murty Mamidanna. Electronic supplementary material The online version of this article (doi:10.1007/s00216-014-7686-y) contains supplementary material, which is available to authorized users. W. G. Mallard (*) Teal Consulting, 7905 Cypress Place, Chevy Chase, MD 20815, USA e-mail: [email protected]
ions that maximize at the same time and examine the time history (shape) of those ions to determine which belong to a particular component. Once a component has been found, the spectrum is extracted along with the retention index, and the spectrum and retention index is compared to the target library spectra and retention indices. The overall method of AMDIS can be characterized as four steps. 1. 2. 3. 4.
Characterization of noise Perception of the component Deconvolution of the component spectrum Comparison of the spectrum and retention data with a target library
The noise characterization not only allows the data to be viewed in terms of signal to noise but also allows for the computation of several parameters that are useful in evaluating the confidence that can be placed in the resulting data. The second step finds each individual component and determines a model shape for the component. The third step extracts a “clean” spectrum from the ion chromatograms using the model shape and, if needed, explicitly subtracting nearby components. Finally, the program takes the extracted spectrum and compares it to library spectra as well as compares the experimental retention index to that of the library compounds.
Noise characterization Noise in any event-counting detector such as an electron multiplier is characterized by a random variation that is proportional to the square root of the signal [2]. The proportionality factor gives a simple measure of the noise. For AMDIS, a “noise factor” is calculated as follows: Nf ¼ ðaverage deviationÞ=signal1=2
W.G. Mallard
The noise factor Nf is calculated from the GC/MS data directly. Segments of the data file that do not show the characteristics of a signal are examined and a value of Nf is found for the dataset. Additional details are given under “Derive Noise Factor” in the Electronic supplementary material (ESM).
Component perception The component perception consists of several steps. A. Correcting the intensities for skewing due to the scanning of the MS (deskewing) B. Selecting ions that maximize in the same time window C. Identifying the extent of possible components and removing baselines D. Finding the time of the maximum (to one tenth of a scan) for each maximizing ion E. Collecting all ions maximizing within ±0.1 scan of the most sharply rising ion F. Defining a model peak for the component The essential element of the perception of components is to find the ions that maximize at the same time. In order to find the ions that maximized at the same time, a correction had to be made for the time between the start of the scan and the time the ion was measured. (A) This deskewing [3, 4] is essential in order to align the mass spectra. In AMDIS, this is done using a parabola defined by the ion intensity scan before the current scan, the current scan and the next scan—each taken at the true time the data were taken, taking into account the scan time. The value of the parabola at the center of the current scan is then taken as the value of the intensity of the ion for the scan (more details in the ESM, “De-skewing the data to align ion profiles”). Once the set of corrected ion time histories are established, (B) the ions that are maximizing are found and the point at which the ion either goes to zero or goes back up after decreasing from the maximum is used as the a definition of the extent of the component (C), at the same time local baselines are removed from each maximizing ion. Only those ion maxima where the size of the maximum is above a predefined level (in terms of noise units) are retained. For all retained maxima, the location of the exact maximum is found (D) by fitting a parabola through the maximum intensity and one scan on either side of the maximum. The peak of this parabola is taken as the retention time of the ion and the abundances are calculated for successive and preceding scans at exactly one scan time from the calculated maximum [5]. Each of the ions that is maximizing is characterized by a “sharpness” value and the sharpness values of ions maximizing in each 0.1 scan bin are summed. A component is perceived from the local maximization of the sharpness sums in a given bin. Then, if there are no other bins with a greater total
sharpness value within a scan range inversely proportional to the total value of the sharpness in the central bin, the component is accepted (E). Finally, a model peak is created from the shape of all of the ions that have sharpness values within 75 % of the ion with maximum sharpness (F). The model peak M(n) is the extracted time history of the component at scan n. By using a set of ions rather than only the sharpest single ion for the model, small peaks that do not have a single dominant ion are better estimated, and noise from a single ion is reduced by the summing of other ions. Additional details are given in the ESM under “Deskewing,” “Defining the retention time of an ion,” and “Defining the component.”
Deconvolution of the component spectrum The spectrum for each component is derived from the model peak created by having the ion chromatogram for each m/z value individual fit to the model profile over the range of the model. AðnÞ ¼ a þ b n þ c M ðnÞ where A(n) is the abundance at scan n and M(n) is the model profile intensity at scan n. The constants a, b, and c are derived from the least-squares fitting of the equation over the range of scans defined by the extent of the model. The terms a and b describe the linear baseline and are not used in the spectral extraction. The abundance for each value of m/z is c×M(nmax), where nmax is the scan number of the maximum model peak abundance. For complex chromatograms, a single model peak may not be sufficient to remove all extraneous ions from the spectrum. In these cases, adjacent model peaks may be explicitly subtracted to get the final spectrum (details in the ESM under “Explicit subtraction of adjacent peaks”). In complex chromatograms, the individual component is interfered with by nearby partially overlapping components. In such cases, the spectra extracted by the least-squares fit above will extract ion abundances for the component that, due to their different shapes, would be judged to arise from the other components. In order to identify the ions that were affected by these interferences, a calculation of the shape of the model relative to the shape of the ion profile in the model window is used. An overlap function, Fm, is calculated for each mass m. X Fm ¼ jI ðnÞ−M ðnÞj summed over n in model peak where I(n) is the intensity of m/z=m at scan n; both I(n) and M(n) have been normalized to ΣI(n)=ΣM(n)=1. For a perfect match, Fm =0; for no overlap, Fm =1. A value of Fm >0.2 caused an ion to be flagged, and a value of Fm >0.6 caused the ion to be rejected from the deconvoluted spectrum. The ions in the spectrum that were flagged are specially handled
AMDIS in the Chemical Weapons Convention
when the deconvoluted spectrum is compared to the target library (below).
Comparison of the spectrum and retention data with a target library The comparison of the extracted spectrum to the target library spectra is done by the calculation of a match factor (MF) using a normalized dot product of the square root of the mass times the intensity for the unknown and library spectra [6]. hX MF ¼ 100 X
1 2
i2
m ðI u I l ÞÞ X m Il m Iu
where the summation is for m=m/z for all ions in either the unknown or the library spectrum; Iu and Il are the intensity at mass m/z for the unknown and library spectra. respectively. The match factor (MF) as shown is the forward match. A reverse match factor containing only the library spectrum data in the denominator (which does not penalize the match for ions that are present in the unknown but not in the library) is also calculated. AMDIS uses a combination of the normal forward match (75 %) and the reverse match (25 %) to get a net match factor. Ions in the deconvoluted spectrum that have been flagged as uncertain, as noted above, are treated differently. Specifically, a match factor is calculated using the flagged ion in the unknown spectrum only if the library spectrum contains that ion. No penalty is assessed for the flagged ion if it does not match the library spectrum. In complex data files with high levels of noise and multiple overlapping components, the ability to use the uncertain peaks in this way significantly increases the ability of AMDIS to identify a target compound. For a perfect match, the value of the match factor is 100; AMDIS will report all matches with a match factor greater than a user-specified value (often 80). There are a number of other corrections to the final match factor that take into account spectra with very few ion peaks, spectra with a single or even two very dominant ion peaks and many very small ion peaks. In addition, corrections are for components that had to have an explicit adjacent peak deconvolution, components that represented a very small portion of the total ion current at the central scan of the deconvoluted model as well as other minor corrections [1] (see ESM under “Corrections to AMDIS match factor” for additional detail). Finally, AMDIS can make use of the retention index measured in the data file. The retention index is calculated either from internal standards in the data file or more commonly from a GC/MS data run under the same chromatographic conditions on a set of linear hydrocarbons (typically C8–C24 for chemical weapon-related analysis) which is used to create
a calibration table that is used by AMDIS to interpolate to find retention index values from the retention time of a component. In AMDIS, the retention index mismatch is used to reduce the net match factor calculated by the spectral match. As will be discussed below, this is critical for the identification of many chemical weapon-related compounds.
Analysis issues specific to the Chemical Weapons Convention Spectral similarity and false positive results Chemical analysis under the CWC includes not only the wellknown chemical weapon compounds, their decomposition products, and precursors but also a far larger range of chemicals that are expected to have properties that would also make them possible chemical agents (see ESM for details of the chemicals covered by the CWC). For analysis to be done using GC/MS and AMDIS, it is necessary for there to be a library of spectra and retention indices of these chemicals so that the spectra extracted by AMDIS can be compared to these spectra. The current version of the database used for this—the OPCW Central Analytical Database—contains over 4,900 mass spectra for over 3,700 distinct chemicals and retention index data for over 3,500 distinct chemicals. The goal of the database is to cover the entire scope of the scheduled chemical weapons, but it also creates some distinct problems. For several classes of chemical agents, the spectral signature for the individual chemicals is very similar across the class. Since the criterion for identification in AMDIS is a preset level of matching between the target spectrum and the experimental data, there can be many target compounds that give match factors that are high enough to be viewed as an identification; these are false positive results, in the strictest sense of the term, but do not represent the identification of a non-scheduled or non-relevant chemical as a CWC-related chemical. In Fig. 1, the spectra and fragmentation patterns for two types of scheduled chemicals are illustrated. For phosphonofluoridate (schedule 1A01), the dominant ion in the spectrum almost always comes from the substituted protonated acid (in Fig. 1, shown for methyl phosphonofluoridate). The 99 (or the corresponding 113 and 127 for the ethyl and the propyl and isopropyl compounds) are the thermodynamically most stable ions and, as a result, often dominate the spectrum. The side chains often do not contribute much to the spectrum and it can be difficult to determine specifically what compound gave rise to the spectrum. If the spectra of these compounds are compared to one another in simple spectral library searches, schedule 1A01 chemicals commonly have three or four distinct compounds, and in some cases, there are as many as 20 compounds that will have a match factor >80 (the cutoff for identification in AMDIS
W.G. Mallard Fig. 1 Spectra for schedule 1A01 methylphophonofluoridate (upper) showing major ions and the fragmentation route to the most common ion. Note that the corresponding ethyl and propyl or isopropyl phosphonofluoridates will have shifts in the major ion reflecting the P-alkyl group (113 and 127, respectively). The lower spectrum is for VX as a representative of the schedule 1A03 dialkylaminoethyl alkylphosphonotiolates; major characteristic ions are identified. Again, note that a single ion dominates the spectrum; however, in this case, the ion does not carry an indication that the molecule contains phosphorus. Only the smaller (and not always present) ion at 79 as well as some much smaller high-mass ions show an indication of the phosphorus content of the starting molecule
when used for on-site inspections). Similar levels of spectral similarity are present in the phosphonate esters and diesters that make up the bulk of the schedule 2B04 chemicals, although the methylphosphonates are slightly less similar than the ethyl, propyl, and isopropyl phosphonates. Note that for all of the schedule 1A01 and 2B04 compounds, the other possible misidentifications will be in the same P-alkyl family. The problem of spectral similarity is most acute for the dialkylaminoethyl alkylphosphonotiolates (schedule 1A03), where the spectra are almost completely dominated by the dialkylaminoethyl portion of the molecule and there is almost no information about the phosphorus linkage. As a result, the misidentifications are not from within the same P-alkyl family, but from within the same N,N-dialkyl family. These compounds—often referred to as V compounds—have such strong spectral similarity that comparing the library spectra
typically gives as many as 40–60 different compounds with mass spectral matches >90 (on the basis of 100) and well over 100 for matches >80. In addition, because the spectra are so dominated by the N,N-dialkyl fragment, there is a greater risk of these molecules having spectral matches to nonphosphorus-containing compounds. Schedule 1A02 chemicals (phosphoramidocyanidates) also show a great deal of similarity, although not as severe as schedule 1A03 chemicals. Typically, there will be five to ten misidentifications for a spectrum from this set, but the highest match factor is, while typically well above 80, for the most part not above 90. In the case of schedule 1A02, as well as the cases for schedules 1A01, 1A03, and 2B04, the fact that the schedules cover an extensive set of chemically and spectrally similar compounds makes these schedules very prone to these misidentifications.
AMDIS in the Chemical Weapons Convention
The discussion above is related just to the comparison of the spectra from the library to one another. Because the identification of the compound by AMDIS depends upon the interference of the spectra by other components in the mixture, it is useful to look at a set of data with some of the scheduled chemicals at various concentration levels in complex matrices to see how many misidentifications do occur. The spectra extracted from the matrix will not be as clean as that from the library and are likely to have interferences due to undeconvolved overlaps that may put spurious ions into the spectrum or make it impossible to extract an ion for the spectrum. At the same time, the role of retention index filtering in removing or reducing the misidentifications will be examined.
Retention index filtering The GC/MS experiment provides two distinct forms of data, the mass spectrum—which is constant over many different types of instruments, but, as noted above, can have high levels of similarity for CWC-related chemicals—and the retention time—which varies depending upon the nature of the column and the conditions of the chromatogram, but measures an entirely independent parameter for the chemical. While the retention time varies greatly, the retention index (RI, which is the time expressed as a fraction of the time that hydrocarbons would elute both before and after the time of the compound) does not vary as much—changing column conditions change the elution time of both the hydrocarbon and the target. In addition, by standardizing the conditions for the chromatographic analysis, the retention indices can become very stable. Since spectral similarity is not related to the retention index, the use of both can often either resolve or at least reduce the misidentifications due to the spectral similarity alone. While spectral similarity is acute for chemical weapon compounds, it also occurs in other classes of spectrally similar chemicals (for example, terpenes, TMS derivatives, unsaturated fatty acids, and some steroids). The retention index is used in AMDIS only as a penalty to the match factor calculated by the mass spectral comparison; that is to say, there is no way for the retention information to increase the spectral match factor, only to diminish it. The penalty is defined as jRI ðExperimentalÞ−RI ðTargetÞj Penalty ¼ −1 RI Window Penalty per RI Window If |RI(Experimental)−RI(Target)|>RI Window; otherwise, Penalty=0. RI(Target) is the retention index of the compound
identified by mass spectral matching. Thus, there is no penalty if the experimental RI is within one window of the library value, and beyond the window, the penalty is proportional to the Penalty per RI Window. The penalty per window is set at 1, 2, 5, 10, 20 or 1,000, with 1,000 being the equivalent of the sharp cut used in some pesticide protocols. The default value is 10, which means that if the RI value of the extracted component which is identified a specific target compound minus the RI value of that target compound is more than three times the RI Window (since the first RI Window is without penalty), then a penalty of 20 (two times the penalty per window) is applied to the mass spectral match factor. The net match factor—which takes into account the spectral match factor minus the RI penalty—cannot be greater than 80 since a perfect match factor is 100. As a result, the compound will not be reported as a match within the framework of a treaty inspection. There is no set value for the RI Window. Small windows with fairly low penalty per window values can distinguish between closely eluting compounds that are spectrally similar. For example, a window of 3 and a penalty per window of 2 would give a penalty of −2 at a RI difference of 5, while a penalty of 10 per window and a window of 10 would not give any penalty at the same RI difference. However, at a RI difference of 30, the small window case would have a penalty of 18 and the large window a penalty of 20. The table below illustrates the problem of spectral similarity for an analysis using the same background interference (an extract of a paint sample) for different target compounds—two from schedule 1A01 and two from schedule 1A03. The AMDIS analysis was done both with and without a retention index penalty. An identification threshold was set to be a net match factor of 80 or greater. When the RI penalty was applied, the retention index window was set to 20 and the penalty level was set to 5 per window. Component at 0.5 ppm
Sarin (1A01) sec-Butyl isopropylphosphonofluoridate (1A01) VX (1A03) Isopropyl S-2-diisopropylaminoethyl isopropylphosphonothiolate (1A03)
Identifications with RI penalty MF ≥80 1
Identifications without RI penalty MF ≥80 4
2
13
2 4
13 51
Similar experiments were run at both higher and lower concentrations. For a complex mixture such as this, it is possible that the best match in terms of the mass spectra is not the best match in terms of the RI value. For example, at
W.G. Mallard
lower concentrations, isopropyl S-2-diisopropylaminoethyl isopropylphosphonothiolate is the third best match for the mass spectrum when the RI values were not used, but the two hits that have better mass spectral matches have RI differences of 92 and 134. These were eliminated by the use of RI filtering (see also “Spectral similarity and RI filtering” in the ESM.) The results shown are consistent with the calculations done with only the library spectra. 1A01 chemicals show a small number of compounds that could be misidentifications and schedule 1A03 chemicals show a far larger number of possible misidentifications. Note that, except for the Sarin data, none of the target compounds was uniquely identified. However, in all cases where RI filtering was used, the correct identification was the one with the highest net match factor. It should be noted that many of the chemicals covered under the CWC can be uniquely identified by just the mass spectra; for example, the sulfur mustards, the nitrogen mustards, and the Lewisites all have very distinctive mass spectra, but even for these chemicals, confidence in the identification is increased by the agreement between the measured and the library retention index.
Strongly overlapped components One of the most severe challenges in any analysis is the separation of the components in a complex mixture so that each can be analyzed by itself. The separation of the components with modern fast mass spectrometers is almost entirely a function of the chromatography. The elution time of a component can be measured in terms of time for the full width at half-maximum (FWHM) of the chromatographic peak. AMDIS is typically capable of separating two components if the peak time of the two components is separated by at least 0.25*FWHM.
Fig. 2 Near-exact co-elutions of two components: 2-methylcyclohexyl methylphosphonofluoridate (2MCH-MPF) at 9.7051 min and Dichlorvos at 9.7152 min. The dots are the actual scan times which are 0.015 min apart. The concentration ratio of the two chemicals is 1:20. The figure on the left shows the experimental TIC and the calculated model for the two
Peak separations shorter than this may still be possible if the peak shapes are different enough. For components with a high degree of similarity in the spectra, it may be necessary to have slightly greater separation. As discussed above, near-overlapping components can give rise to ion peaks that cannot be confidently assigned to the component or rejected from it. AMDIS handles these by flagging them and then calculating the match factor both with and without the ions. An example of this is given with data for two very closely eluting compounds. Here, there are two compounds—2-methylcyclohexyl methylphosphonofluoridate (2MCH-MPF), which is a schedule 1A01 chemical, and Dichlovos, a pesticide—which are present in a 1:20 ratio and have maxima at retention times 9.7051 and 9.7152, respectively. The data for the total ion chromatogram (TIC) and the extracted models for the two components are shown on the left in Fig. 2. The scaled data for a single model (that of the first component) and one ion are shown on the right. The areas under the two curves in the righthand figure are the same, and the comparison of the shapes of the two curves shows that they may not represent the same physical eluent. The program flags the m/z109 peak for the component. The deconvoluted spectrum extracted for the first component as shown in AMDIS has both certain peaks (solid lines) and uncertain peaks (such as the m/z109, shown as dashed lines); the library spectrum for the target compound is shown in red (shown on the left in Fig. 3). Note that the apparent match between the library spectrum and the experimental data is very poor in the data on the left; in fact, if the spectrum (both solid and dashed peaks) is searched, the first match will be Dichlorvos, not 2MCH-MPF. However, the AMDIS match factor for the 2MCH-MPF target compound is 88. To get this kind of a match factor, AMDIS takes into account the flagged (dashed) peaks and does a match factor calculation both with and without the
components. The figure on the right shows the model for the 2-MCHMPF along with the m/z109 peak (which arises from the Dichlorovos). Note the mismatch in the shape of the individual ion compared to the model of the component
AMDIS in the Chemical Weapons Convention
Fig. 3 The graph on the left shows the spectrum of the component in black with both the flagged (dashed) and unflagged (solid) peaks as well as the library spectrum of 2-methylcyclohexyl methylphosphonofluoridate (2MCH-MPF) in red. Note that the unflagged peaks are so small that they are difficult to see. The total ion current associated with the unflagged peaks
in this spectrum is only 11 %. The spectra on the right are again of the library (red) and the component (black), but with only the unflagged peaks from the deconvoluted spectrum. Note the good agreement between the two spectra here
flagged peaks (shown on the right of the figure). In this graph, it is clear that the extracted spectrum and the library spectrum are very similar. The high level of the total ion current of the extracted spectrum that is due to the uncertain peaks does penalize the match factor somewhat—the raw match factor for just the certain peaks is over 95. Note that the use of uncertain peaks varies with each target spectrum. If a target spectrum has a peak that AMDIS has flagged as uncertain, the peak is counted toward the match. This means that AMDIS is less likely to miss an ion that does belong to a particular component and thus not count it toward the match factor and at the same time is less likely to accept ions that do not match the library and thus not penalize the match due to the interference. In both cases, the use or nonuse of uncertain peaks increases the chance that AMDIS will identify a component. A second common problem with overlapped data is that a set of ions are not extracted into the spectrum, in some sense the exact reverse of what was discussed above. An example of this is shown in Fig. 4. The ions maximizing at about 16.57 min in the graph (m/z 357 and 299) are from 2glycerophosphoric acid-4TMS. The base peak in the spectrum is m/z73, and both m/z103 and 147 are about 20 % of the base peak. The extracted spectrum (Fig. 5) does not show the peaks at 73, 103, or 147, along with many other low mass peaks that are not extracted.
at low concentrations. In addition, the method by which AMDIS creates model peaks allows it to find even very small peaks and look at the ions that rise and fall with those peaks. The resulting spectra are then compared to the target library. For most target library searches, (e.g., pesticides, chemical weapons, drugs of abuse), these multiple models greatly increase the probability that AMDIS will identify the target molecule. It is impossible to guarantee that the target of 0.1 ppm of any target compound in any matrix can be reached—if the target compound has a mass spectrum very similar to the spectra from the matrix that elute near the target compound or if the matrix is very complex, the goal may be essentially impossible. However, for a wide range of matrices, sensitivity at or near this level can be achieved. To achieve this sensitivity, AMDIS will look at a given peak in a number of ways, and if the resulting spectra differ, then all of these will be reported. It is not unusual in a complex data file wherein for what appears to an analyst to be a single
Sensitivity and filtering The original design of AMDIS was focused on ensuring that if a target compound was in a complex matrix, it would be reported if its concentration was above 0.1 ppm. As discussed above, the use of uncertain peaks gives AMDIS the ability to find highly overlapped components
Fig. 4 The component that is peaking at 16.57 can be seen in the maxima of the m/z299 and 357 ions, but not at all in the TIC. Ions that should be present in the deconvoluted spectrum (103 and 147 are illustrated here) do not show any peak since both are in larger components on either side of the target component. The result will be a spectrum that is missing these ions
W.G. Mallard Fig. 5 The top spectrum (red) is that extracted from the data file. Notice that almost all of the lowmass ions are not extracted. On either side of the extracted component, there are components that have m/z73, 103, 147 and most of the other ions that belong to the target compound. Identification can be made given the high-mass ions and the retention index, but this example shows the kinds of limitations that occur in highly congested data files
peak will be divided by AMDIS into multiple components. This procedure is essential in finding cases where an apparent single peak in the total ion chromatogram is actually the sum of several components. However, it is also the case that noise in the data can produce artificial peaks that do not correspond to true components. It is possible to reduce the number of components by lowering the sensitivity of AMDIS, but in doing so, it is possible to lose low-concentration target peaks. A better way is to find parameters that come out of the analysis that show that the resulting component may not be a complete deconvolution. To accomplish this, a large number of data files were examined to determine what characteristics could best be used to eliminate these possibly spurious peaks. There are a number of statistical measures that AMDIS creates for each component it extracts; these include measures of the purity of the component (i.e., the fraction of the TIC accounted for by a specific component), the total integrated signal from the component, and various measures that indicate the quality of the deconvolution. A large number of data files with significant degrees of interference from the background, adjacent peaks, broad peaks covering up components under them, as well as other problems in GC/MS files and individual components were examined to find data that the analyst would not think was a good spectrum. From this, a number of filters were developed that allowed peaks to be removed. The filters that were most effective (and were subsequently incorporated into AMDIS) were 1. The number of model ions: As noted above, a model is created from all of the ions which have “sharpness” values equal to 75 % of the ion with the maximum “sharpness”. As the number of such ions increases, it means that more ions can be extracted that appear to have nearly the same time history. As this number increases, it increases the confidence that a single component is being deconvoluted.
2. The fraction of the extracted intensity that consists of flagged peaks: For highly overlapped components, it is often the case that a substantial fraction of the extracted ion peak intensity has been flagged. In many cases, small noise peaks can be picked up for models, and substantial fractions of the extracted ions are then uncertain. As this fraction of flagged peaks decreases, or conversely as the fraction of good peaks increases, the confidence in the identification of a single component increases. 3. The minimum abundance: For any spectrum, the smallest relative to the base peak is a function of the noise level in the data and the size of the base peak. As the minimum abundance increases, the confidence that low-intensity peaks in the spectrum can be extracted decreases. 4. The signal-to-noise ratio (S/N) of a component: AMDIS calculates a total S/N for every component. This value is a function not only of the size of the base peak but also the number of ions that were extracted. Again, as this value increases, so does the confidence in a reliable deconvolution. Each of these factors can be used either separately as a cutoff filter or in combination in a combined filter that allows for some of the thresholds to be missed if others are exceeded. The filters are tools for use in the simplification of data where target analysis is not the sole reason for the AMDIS analysis. The filters are only applied to components that have not been matched by the target search; to do so otherwise might risk losing a match. For example, the spectrum shown in Fig. 3 had a very low fraction of the extracted ion intensity that was not flagged (11 %), and had that parameter been used as an cutoff filter (i.e., not accepting any data with fraction
FEATURE ARTICLE
AMDIS in the Chemical Weapons Convention W. Gary Mallard
# Springer-Verlag Berlin Heidelberg 2014
Introduction The inspection regime under the Chemical Weapons Convention (CWC) requires both the ability to identify low levels of chemical agents, precursors, and decomposition products in complex mixtures while at the same time not exposing the inspection site to the risk of loss of confidential information related to the chemical composition of the sample. The method of choice for such detection of low levels of chemical agents is gas chromatography separation with mass spectral identification (GC/MS), but the usual method of such analysis—an extensive examination of the GC/MS data file for spectra matching the compounds of interest—is both time-consuming, prone to missing low-level components, and difficult to ensure that chemicals not relevant to the CWC are not discovered. To deal with these problems, AMDIS (Automatic Mass spectral Deconvolution and Identification System), a method of automatically finding all of the distinct chemical components in the GC/MS data file and then compare them to a library of spectra of chemicals relevant to the CWC was developed.
Fundamental algorithms of AMDIS The basic concept behind AMDIS [1] is to examine the GC/ MS data file and find all instances of ion maximization; then, for each of these cases where a maximum is found, to find all Published in the topical collection Analysis of Chemicals Relevant to the Chemical Weapons Convention with guest editors Marc-Michael Blum and R. V. S. Murty Mamidanna. Electronic supplementary material The online version of this article (doi:10.1007/s00216-014-7686-y) contains supplementary material, which is available to authorized users. W. G. Mallard (*) Teal Consulting, 7905 Cypress Place, Chevy Chase, MD 20815, USA e-mail: [email protected]
ions that maximize at the same time and examine the time history (shape) of those ions to determine which belong to a particular component. Once a component has been found, the spectrum is extracted along with the retention index, and the spectrum and retention index is compared to the target library spectra and retention indices. The overall method of AMDIS can be characterized as four steps. 1. 2. 3. 4.
Characterization of noise Perception of the component Deconvolution of the component spectrum Comparison of the spectrum and retention data with a target library
The noise characterization not only allows the data to be viewed in terms of signal to noise but also allows for the computation of several parameters that are useful in evaluating the confidence that can be placed in the resulting data. The second step finds each individual component and determines a model shape for the component. The third step extracts a “clean” spectrum from the ion chromatograms using the model shape and, if needed, explicitly subtracting nearby components. Finally, the program takes the extracted spectrum and compares it to library spectra as well as compares the experimental retention index to that of the library compounds.
Noise characterization Noise in any event-counting detector such as an electron multiplier is characterized by a random variation that is proportional to the square root of the signal [2]. The proportionality factor gives a simple measure of the noise. For AMDIS, a “noise factor” is calculated as follows: Nf ¼ ðaverage deviationÞ=signal1=2
W.G. Mallard
The noise factor Nf is calculated from the GC/MS data directly. Segments of the data file that do not show the characteristics of a signal are examined and a value of Nf is found for the dataset. Additional details are given under “Derive Noise Factor” in the Electronic supplementary material (ESM).
Component perception The component perception consists of several steps. A. Correcting the intensities for skewing due to the scanning of the MS (deskewing) B. Selecting ions that maximize in the same time window C. Identifying the extent of possible components and removing baselines D. Finding the time of the maximum (to one tenth of a scan) for each maximizing ion E. Collecting all ions maximizing within ±0.1 scan of the most sharply rising ion F. Defining a model peak for the component The essential element of the perception of components is to find the ions that maximize at the same time. In order to find the ions that maximized at the same time, a correction had to be made for the time between the start of the scan and the time the ion was measured. (A) This deskewing [3, 4] is essential in order to align the mass spectra. In AMDIS, this is done using a parabola defined by the ion intensity scan before the current scan, the current scan and the next scan—each taken at the true time the data were taken, taking into account the scan time. The value of the parabola at the center of the current scan is then taken as the value of the intensity of the ion for the scan (more details in the ESM, “De-skewing the data to align ion profiles”). Once the set of corrected ion time histories are established, (B) the ions that are maximizing are found and the point at which the ion either goes to zero or goes back up after decreasing from the maximum is used as the a definition of the extent of the component (C), at the same time local baselines are removed from each maximizing ion. Only those ion maxima where the size of the maximum is above a predefined level (in terms of noise units) are retained. For all retained maxima, the location of the exact maximum is found (D) by fitting a parabola through the maximum intensity and one scan on either side of the maximum. The peak of this parabola is taken as the retention time of the ion and the abundances are calculated for successive and preceding scans at exactly one scan time from the calculated maximum [5]. Each of the ions that is maximizing is characterized by a “sharpness” value and the sharpness values of ions maximizing in each 0.1 scan bin are summed. A component is perceived from the local maximization of the sharpness sums in a given bin. Then, if there are no other bins with a greater total
sharpness value within a scan range inversely proportional to the total value of the sharpness in the central bin, the component is accepted (E). Finally, a model peak is created from the shape of all of the ions that have sharpness values within 75 % of the ion with maximum sharpness (F). The model peak M(n) is the extracted time history of the component at scan n. By using a set of ions rather than only the sharpest single ion for the model, small peaks that do not have a single dominant ion are better estimated, and noise from a single ion is reduced by the summing of other ions. Additional details are given in the ESM under “Deskewing,” “Defining the retention time of an ion,” and “Defining the component.”
Deconvolution of the component spectrum The spectrum for each component is derived from the model peak created by having the ion chromatogram for each m/z value individual fit to the model profile over the range of the model. AðnÞ ¼ a þ b n þ c M ðnÞ where A(n) is the abundance at scan n and M(n) is the model profile intensity at scan n. The constants a, b, and c are derived from the least-squares fitting of the equation over the range of scans defined by the extent of the model. The terms a and b describe the linear baseline and are not used in the spectral extraction. The abundance for each value of m/z is c×M(nmax), where nmax is the scan number of the maximum model peak abundance. For complex chromatograms, a single model peak may not be sufficient to remove all extraneous ions from the spectrum. In these cases, adjacent model peaks may be explicitly subtracted to get the final spectrum (details in the ESM under “Explicit subtraction of adjacent peaks”). In complex chromatograms, the individual component is interfered with by nearby partially overlapping components. In such cases, the spectra extracted by the least-squares fit above will extract ion abundances for the component that, due to their different shapes, would be judged to arise from the other components. In order to identify the ions that were affected by these interferences, a calculation of the shape of the model relative to the shape of the ion profile in the model window is used. An overlap function, Fm, is calculated for each mass m. X Fm ¼ jI ðnÞ−M ðnÞj summed over n in model peak where I(n) is the intensity of m/z=m at scan n; both I(n) and M(n) have been normalized to ΣI(n)=ΣM(n)=1. For a perfect match, Fm =0; for no overlap, Fm =1. A value of Fm >0.2 caused an ion to be flagged, and a value of Fm >0.6 caused the ion to be rejected from the deconvoluted spectrum. The ions in the spectrum that were flagged are specially handled
AMDIS in the Chemical Weapons Convention
when the deconvoluted spectrum is compared to the target library (below).
Comparison of the spectrum and retention data with a target library The comparison of the extracted spectrum to the target library spectra is done by the calculation of a match factor (MF) using a normalized dot product of the square root of the mass times the intensity for the unknown and library spectra [6]. hX MF ¼ 100 X
1 2
i2
m ðI u I l ÞÞ X m Il m Iu
where the summation is for m=m/z for all ions in either the unknown or the library spectrum; Iu and Il are the intensity at mass m/z for the unknown and library spectra. respectively. The match factor (MF) as shown is the forward match. A reverse match factor containing only the library spectrum data in the denominator (which does not penalize the match for ions that are present in the unknown but not in the library) is also calculated. AMDIS uses a combination of the normal forward match (75 %) and the reverse match (25 %) to get a net match factor. Ions in the deconvoluted spectrum that have been flagged as uncertain, as noted above, are treated differently. Specifically, a match factor is calculated using the flagged ion in the unknown spectrum only if the library spectrum contains that ion. No penalty is assessed for the flagged ion if it does not match the library spectrum. In complex data files with high levels of noise and multiple overlapping components, the ability to use the uncertain peaks in this way significantly increases the ability of AMDIS to identify a target compound. For a perfect match, the value of the match factor is 100; AMDIS will report all matches with a match factor greater than a user-specified value (often 80). There are a number of other corrections to the final match factor that take into account spectra with very few ion peaks, spectra with a single or even two very dominant ion peaks and many very small ion peaks. In addition, corrections are for components that had to have an explicit adjacent peak deconvolution, components that represented a very small portion of the total ion current at the central scan of the deconvoluted model as well as other minor corrections [1] (see ESM under “Corrections to AMDIS match factor” for additional detail). Finally, AMDIS can make use of the retention index measured in the data file. The retention index is calculated either from internal standards in the data file or more commonly from a GC/MS data run under the same chromatographic conditions on a set of linear hydrocarbons (typically C8–C24 for chemical weapon-related analysis) which is used to create
a calibration table that is used by AMDIS to interpolate to find retention index values from the retention time of a component. In AMDIS, the retention index mismatch is used to reduce the net match factor calculated by the spectral match. As will be discussed below, this is critical for the identification of many chemical weapon-related compounds.
Analysis issues specific to the Chemical Weapons Convention Spectral similarity and false positive results Chemical analysis under the CWC includes not only the wellknown chemical weapon compounds, their decomposition products, and precursors but also a far larger range of chemicals that are expected to have properties that would also make them possible chemical agents (see ESM for details of the chemicals covered by the CWC). For analysis to be done using GC/MS and AMDIS, it is necessary for there to be a library of spectra and retention indices of these chemicals so that the spectra extracted by AMDIS can be compared to these spectra. The current version of the database used for this—the OPCW Central Analytical Database—contains over 4,900 mass spectra for over 3,700 distinct chemicals and retention index data for over 3,500 distinct chemicals. The goal of the database is to cover the entire scope of the scheduled chemical weapons, but it also creates some distinct problems. For several classes of chemical agents, the spectral signature for the individual chemicals is very similar across the class. Since the criterion for identification in AMDIS is a preset level of matching between the target spectrum and the experimental data, there can be many target compounds that give match factors that are high enough to be viewed as an identification; these are false positive results, in the strictest sense of the term, but do not represent the identification of a non-scheduled or non-relevant chemical as a CWC-related chemical. In Fig. 1, the spectra and fragmentation patterns for two types of scheduled chemicals are illustrated. For phosphonofluoridate (schedule 1A01), the dominant ion in the spectrum almost always comes from the substituted protonated acid (in Fig. 1, shown for methyl phosphonofluoridate). The 99 (or the corresponding 113 and 127 for the ethyl and the propyl and isopropyl compounds) are the thermodynamically most stable ions and, as a result, often dominate the spectrum. The side chains often do not contribute much to the spectrum and it can be difficult to determine specifically what compound gave rise to the spectrum. If the spectra of these compounds are compared to one another in simple spectral library searches, schedule 1A01 chemicals commonly have three or four distinct compounds, and in some cases, there are as many as 20 compounds that will have a match factor >80 (the cutoff for identification in AMDIS
W.G. Mallard Fig. 1 Spectra for schedule 1A01 methylphophonofluoridate (upper) showing major ions and the fragmentation route to the most common ion. Note that the corresponding ethyl and propyl or isopropyl phosphonofluoridates will have shifts in the major ion reflecting the P-alkyl group (113 and 127, respectively). The lower spectrum is for VX as a representative of the schedule 1A03 dialkylaminoethyl alkylphosphonotiolates; major characteristic ions are identified. Again, note that a single ion dominates the spectrum; however, in this case, the ion does not carry an indication that the molecule contains phosphorus. Only the smaller (and not always present) ion at 79 as well as some much smaller high-mass ions show an indication of the phosphorus content of the starting molecule
when used for on-site inspections). Similar levels of spectral similarity are present in the phosphonate esters and diesters that make up the bulk of the schedule 2B04 chemicals, although the methylphosphonates are slightly less similar than the ethyl, propyl, and isopropyl phosphonates. Note that for all of the schedule 1A01 and 2B04 compounds, the other possible misidentifications will be in the same P-alkyl family. The problem of spectral similarity is most acute for the dialkylaminoethyl alkylphosphonotiolates (schedule 1A03), where the spectra are almost completely dominated by the dialkylaminoethyl portion of the molecule and there is almost no information about the phosphorus linkage. As a result, the misidentifications are not from within the same P-alkyl family, but from within the same N,N-dialkyl family. These compounds—often referred to as V compounds—have such strong spectral similarity that comparing the library spectra
typically gives as many as 40–60 different compounds with mass spectral matches >90 (on the basis of 100) and well over 100 for matches >80. In addition, because the spectra are so dominated by the N,N-dialkyl fragment, there is a greater risk of these molecules having spectral matches to nonphosphorus-containing compounds. Schedule 1A02 chemicals (phosphoramidocyanidates) also show a great deal of similarity, although not as severe as schedule 1A03 chemicals. Typically, there will be five to ten misidentifications for a spectrum from this set, but the highest match factor is, while typically well above 80, for the most part not above 90. In the case of schedule 1A02, as well as the cases for schedules 1A01, 1A03, and 2B04, the fact that the schedules cover an extensive set of chemically and spectrally similar compounds makes these schedules very prone to these misidentifications.
AMDIS in the Chemical Weapons Convention
The discussion above is related just to the comparison of the spectra from the library to one another. Because the identification of the compound by AMDIS depends upon the interference of the spectra by other components in the mixture, it is useful to look at a set of data with some of the scheduled chemicals at various concentration levels in complex matrices to see how many misidentifications do occur. The spectra extracted from the matrix will not be as clean as that from the library and are likely to have interferences due to undeconvolved overlaps that may put spurious ions into the spectrum or make it impossible to extract an ion for the spectrum. At the same time, the role of retention index filtering in removing or reducing the misidentifications will be examined.
Retention index filtering The GC/MS experiment provides two distinct forms of data, the mass spectrum—which is constant over many different types of instruments, but, as noted above, can have high levels of similarity for CWC-related chemicals—and the retention time—which varies depending upon the nature of the column and the conditions of the chromatogram, but measures an entirely independent parameter for the chemical. While the retention time varies greatly, the retention index (RI, which is the time expressed as a fraction of the time that hydrocarbons would elute both before and after the time of the compound) does not vary as much—changing column conditions change the elution time of both the hydrocarbon and the target. In addition, by standardizing the conditions for the chromatographic analysis, the retention indices can become very stable. Since spectral similarity is not related to the retention index, the use of both can often either resolve or at least reduce the misidentifications due to the spectral similarity alone. While spectral similarity is acute for chemical weapon compounds, it also occurs in other classes of spectrally similar chemicals (for example, terpenes, TMS derivatives, unsaturated fatty acids, and some steroids). The retention index is used in AMDIS only as a penalty to the match factor calculated by the mass spectral comparison; that is to say, there is no way for the retention information to increase the spectral match factor, only to diminish it. The penalty is defined as jRI ðExperimentalÞ−RI ðTargetÞj Penalty ¼ −1 RI Window Penalty per RI Window If |RI(Experimental)−RI(Target)|>RI Window; otherwise, Penalty=0. RI(Target) is the retention index of the compound
identified by mass spectral matching. Thus, there is no penalty if the experimental RI is within one window of the library value, and beyond the window, the penalty is proportional to the Penalty per RI Window. The penalty per window is set at 1, 2, 5, 10, 20 or 1,000, with 1,000 being the equivalent of the sharp cut used in some pesticide protocols. The default value is 10, which means that if the RI value of the extracted component which is identified a specific target compound minus the RI value of that target compound is more than three times the RI Window (since the first RI Window is without penalty), then a penalty of 20 (two times the penalty per window) is applied to the mass spectral match factor. The net match factor—which takes into account the spectral match factor minus the RI penalty—cannot be greater than 80 since a perfect match factor is 100. As a result, the compound will not be reported as a match within the framework of a treaty inspection. There is no set value for the RI Window. Small windows with fairly low penalty per window values can distinguish between closely eluting compounds that are spectrally similar. For example, a window of 3 and a penalty per window of 2 would give a penalty of −2 at a RI difference of 5, while a penalty of 10 per window and a window of 10 would not give any penalty at the same RI difference. However, at a RI difference of 30, the small window case would have a penalty of 18 and the large window a penalty of 20. The table below illustrates the problem of spectral similarity for an analysis using the same background interference (an extract of a paint sample) for different target compounds—two from schedule 1A01 and two from schedule 1A03. The AMDIS analysis was done both with and without a retention index penalty. An identification threshold was set to be a net match factor of 80 or greater. When the RI penalty was applied, the retention index window was set to 20 and the penalty level was set to 5 per window. Component at 0.5 ppm
Sarin (1A01) sec-Butyl isopropylphosphonofluoridate (1A01) VX (1A03) Isopropyl S-2-diisopropylaminoethyl isopropylphosphonothiolate (1A03)
Identifications with RI penalty MF ≥80 1
Identifications without RI penalty MF ≥80 4
2
13
2 4
13 51
Similar experiments were run at both higher and lower concentrations. For a complex mixture such as this, it is possible that the best match in terms of the mass spectra is not the best match in terms of the RI value. For example, at
W.G. Mallard
lower concentrations, isopropyl S-2-diisopropylaminoethyl isopropylphosphonothiolate is the third best match for the mass spectrum when the RI values were not used, but the two hits that have better mass spectral matches have RI differences of 92 and 134. These were eliminated by the use of RI filtering (see also “Spectral similarity and RI filtering” in the ESM.) The results shown are consistent with the calculations done with only the library spectra. 1A01 chemicals show a small number of compounds that could be misidentifications and schedule 1A03 chemicals show a far larger number of possible misidentifications. Note that, except for the Sarin data, none of the target compounds was uniquely identified. However, in all cases where RI filtering was used, the correct identification was the one with the highest net match factor. It should be noted that many of the chemicals covered under the CWC can be uniquely identified by just the mass spectra; for example, the sulfur mustards, the nitrogen mustards, and the Lewisites all have very distinctive mass spectra, but even for these chemicals, confidence in the identification is increased by the agreement between the measured and the library retention index.
Strongly overlapped components One of the most severe challenges in any analysis is the separation of the components in a complex mixture so that each can be analyzed by itself. The separation of the components with modern fast mass spectrometers is almost entirely a function of the chromatography. The elution time of a component can be measured in terms of time for the full width at half-maximum (FWHM) of the chromatographic peak. AMDIS is typically capable of separating two components if the peak time of the two components is separated by at least 0.25*FWHM.
Fig. 2 Near-exact co-elutions of two components: 2-methylcyclohexyl methylphosphonofluoridate (2MCH-MPF) at 9.7051 min and Dichlorvos at 9.7152 min. The dots are the actual scan times which are 0.015 min apart. The concentration ratio of the two chemicals is 1:20. The figure on the left shows the experimental TIC and the calculated model for the two
Peak separations shorter than this may still be possible if the peak shapes are different enough. For components with a high degree of similarity in the spectra, it may be necessary to have slightly greater separation. As discussed above, near-overlapping components can give rise to ion peaks that cannot be confidently assigned to the component or rejected from it. AMDIS handles these by flagging them and then calculating the match factor both with and without the ions. An example of this is given with data for two very closely eluting compounds. Here, there are two compounds—2-methylcyclohexyl methylphosphonofluoridate (2MCH-MPF), which is a schedule 1A01 chemical, and Dichlovos, a pesticide—which are present in a 1:20 ratio and have maxima at retention times 9.7051 and 9.7152, respectively. The data for the total ion chromatogram (TIC) and the extracted models for the two components are shown on the left in Fig. 2. The scaled data for a single model (that of the first component) and one ion are shown on the right. The areas under the two curves in the righthand figure are the same, and the comparison of the shapes of the two curves shows that they may not represent the same physical eluent. The program flags the m/z109 peak for the component. The deconvoluted spectrum extracted for the first component as shown in AMDIS has both certain peaks (solid lines) and uncertain peaks (such as the m/z109, shown as dashed lines); the library spectrum for the target compound is shown in red (shown on the left in Fig. 3). Note that the apparent match between the library spectrum and the experimental data is very poor in the data on the left; in fact, if the spectrum (both solid and dashed peaks) is searched, the first match will be Dichlorvos, not 2MCH-MPF. However, the AMDIS match factor for the 2MCH-MPF target compound is 88. To get this kind of a match factor, AMDIS takes into account the flagged (dashed) peaks and does a match factor calculation both with and without the
components. The figure on the right shows the model for the 2-MCHMPF along with the m/z109 peak (which arises from the Dichlorovos). Note the mismatch in the shape of the individual ion compared to the model of the component
AMDIS in the Chemical Weapons Convention
Fig. 3 The graph on the left shows the spectrum of the component in black with both the flagged (dashed) and unflagged (solid) peaks as well as the library spectrum of 2-methylcyclohexyl methylphosphonofluoridate (2MCH-MPF) in red. Note that the unflagged peaks are so small that they are difficult to see. The total ion current associated with the unflagged peaks
in this spectrum is only 11 %. The spectra on the right are again of the library (red) and the component (black), but with only the unflagged peaks from the deconvoluted spectrum. Note the good agreement between the two spectra here
flagged peaks (shown on the right of the figure). In this graph, it is clear that the extracted spectrum and the library spectrum are very similar. The high level of the total ion current of the extracted spectrum that is due to the uncertain peaks does penalize the match factor somewhat—the raw match factor for just the certain peaks is over 95. Note that the use of uncertain peaks varies with each target spectrum. If a target spectrum has a peak that AMDIS has flagged as uncertain, the peak is counted toward the match. This means that AMDIS is less likely to miss an ion that does belong to a particular component and thus not count it toward the match factor and at the same time is less likely to accept ions that do not match the library and thus not penalize the match due to the interference. In both cases, the use or nonuse of uncertain peaks increases the chance that AMDIS will identify a component. A second common problem with overlapped data is that a set of ions are not extracted into the spectrum, in some sense the exact reverse of what was discussed above. An example of this is shown in Fig. 4. The ions maximizing at about 16.57 min in the graph (m/z 357 and 299) are from 2glycerophosphoric acid-4TMS. The base peak in the spectrum is m/z73, and both m/z103 and 147 are about 20 % of the base peak. The extracted spectrum (Fig. 5) does not show the peaks at 73, 103, or 147, along with many other low mass peaks that are not extracted.
at low concentrations. In addition, the method by which AMDIS creates model peaks allows it to find even very small peaks and look at the ions that rise and fall with those peaks. The resulting spectra are then compared to the target library. For most target library searches, (e.g., pesticides, chemical weapons, drugs of abuse), these multiple models greatly increase the probability that AMDIS will identify the target molecule. It is impossible to guarantee that the target of 0.1 ppm of any target compound in any matrix can be reached—if the target compound has a mass spectrum very similar to the spectra from the matrix that elute near the target compound or if the matrix is very complex, the goal may be essentially impossible. However, for a wide range of matrices, sensitivity at or near this level can be achieved. To achieve this sensitivity, AMDIS will look at a given peak in a number of ways, and if the resulting spectra differ, then all of these will be reported. It is not unusual in a complex data file wherein for what appears to an analyst to be a single
Sensitivity and filtering The original design of AMDIS was focused on ensuring that if a target compound was in a complex matrix, it would be reported if its concentration was above 0.1 ppm. As discussed above, the use of uncertain peaks gives AMDIS the ability to find highly overlapped components
Fig. 4 The component that is peaking at 16.57 can be seen in the maxima of the m/z299 and 357 ions, but not at all in the TIC. Ions that should be present in the deconvoluted spectrum (103 and 147 are illustrated here) do not show any peak since both are in larger components on either side of the target component. The result will be a spectrum that is missing these ions
W.G. Mallard Fig. 5 The top spectrum (red) is that extracted from the data file. Notice that almost all of the lowmass ions are not extracted. On either side of the extracted component, there are components that have m/z73, 103, 147 and most of the other ions that belong to the target compound. Identification can be made given the high-mass ions and the retention index, but this example shows the kinds of limitations that occur in highly congested data files
peak will be divided by AMDIS into multiple components. This procedure is essential in finding cases where an apparent single peak in the total ion chromatogram is actually the sum of several components. However, it is also the case that noise in the data can produce artificial peaks that do not correspond to true components. It is possible to reduce the number of components by lowering the sensitivity of AMDIS, but in doing so, it is possible to lose low-concentration target peaks. A better way is to find parameters that come out of the analysis that show that the resulting component may not be a complete deconvolution. To accomplish this, a large number of data files were examined to determine what characteristics could best be used to eliminate these possibly spurious peaks. There are a number of statistical measures that AMDIS creates for each component it extracts; these include measures of the purity of the component (i.e., the fraction of the TIC accounted for by a specific component), the total integrated signal from the component, and various measures that indicate the quality of the deconvolution. A large number of data files with significant degrees of interference from the background, adjacent peaks, broad peaks covering up components under them, as well as other problems in GC/MS files and individual components were examined to find data that the analyst would not think was a good spectrum. From this, a number of filters were developed that allowed peaks to be removed. The filters that were most effective (and were subsequently incorporated into AMDIS) were 1. The number of model ions: As noted above, a model is created from all of the ions which have “sharpness” values equal to 75 % of the ion with the maximum “sharpness”. As the number of such ions increases, it means that more ions can be extracted that appear to have nearly the same time history. As this number increases, it increases the confidence that a single component is being deconvoluted.
2. The fraction of the extracted intensity that consists of flagged peaks: For highly overlapped components, it is often the case that a substantial fraction of the extracted ion peak intensity has been flagged. In many cases, small noise peaks can be picked up for models, and substantial fractions of the extracted ions are then uncertain. As this fraction of flagged peaks decreases, or conversely as the fraction of good peaks increases, the confidence in the identification of a single component increases. 3. The minimum abundance: For any spectrum, the smallest relative to the base peak is a function of the noise level in the data and the size of the base peak. As the minimum abundance increases, the confidence that low-intensity peaks in the spectrum can be extracted decreases. 4. The signal-to-noise ratio (S/N) of a component: AMDIS calculates a total S/N for every component. This value is a function not only of the size of the base peak but also the number of ions that were extracted. Again, as this value increases, so does the confidence in a reliable deconvolution. Each of these factors can be used either separately as a cutoff filter or in combination in a combined filter that allows for some of the thresholds to be missed if others are exceeded. The filters are tools for use in the simplification of data where target analysis is not the sole reason for the AMDIS analysis. The filters are only applied to components that have not been matched by the target search; to do so otherwise might risk losing a match. For example, the spectrum shown in Fig. 3 had a very low fraction of the extracted ion intensity that was not flagged (11 %), and had that parameter been used as an cutoff filter (i.e., not accepting any data with fraction
