Automating Data Analysis for Two-Dimensional Gas Chromatography/Time-of-Flight Mass Spectrometry Non-Targeted Analysis of Comparative Samples

Abstract

Non-targeted analysis of environmental samples, using comprehensive two-dimensional gas chromatography coupled with time-of-flight mass spectrometry (GC×GC/ToF-MS), poses significant data analysis challenges due to the large number of possible analytes. Non-targeted data analysis of complex mixtures is prone to human bias and is laborious, particularly for comparative environmental samples such as contaminated soil pre- and post-bioremediation. To address this research bottleneck, we developed OCTpy, a Python^™ script that acts as a data reduction filter to automate GC×GC/ToF-MS data analysis from LECO^® ChromaTOF^® software and facilitates selection of analytes of interest based on peak area comparison between comparative samples. We used data from polycyclic aromatic hydrocarbon (PAH) contaminated soil, pre- and post-bioremediation, to assess the effectiveness of OCTpy in facilitating the selection of analytes that have formed or degraded following treatment. Using datasets from the soil extracts pre- and post-bioremediation, OCTpy selected, on average, 18% of the initial suggested analytes generated by the LECO^® ChromaTOF^® software Statistical Compare feature. Based on this list, 63–100% of the candidate analytes identified by a highly trained individual were also selected by OCTpy. This process was accomplished in several minutes per sample, whereas manual data analysis took several hours per sample. OCTpy automates the analysis of complex mixtures of comparative samples, reduces the potential for human error during heavy data handling and decreases data analysis time by at least tenfold.

1. Introduction

Comprehensive two-dimensional gas chromatography coupled with time-of-flight mass spectrometry (GC×GC/ToF-MS) is a versatile tool capable of elucidating organic analytes in complex environmental mixtures.[1–3] GC×GC/ToF-MS is also useful in non-targeted analysis[4–6] to help narrow down or identify analytes of interest, due to its higher peak capacity, sensitivity, and selectivity.[7–11] However, GC×GC/ToF-MS generates a large amount of data, which can be a major bottleneck in data analysis. Handling large volumes of data is not only time-consuming, particularly in non-targeted analysis, but is also prone to inter-laboratory discrepancies due to human error, regardless of the sample matrix.[3,7,12,13] For these reasons, an automated approach for GC×GC/ToF-MS data analysis is desired.

Historically, Microsoft^® VBScripts,[1] C++,[2] MathWorks^® MATLAB[14], and Guineu[15] have been the programming language or computer programs of choice for the development of data analysis scripts. Python^™ has not been applied for the analysis of GC×GC/ToF-MS data despite its widespread use,[16–18] including for GC/MS data analysis.[19–21] Furthermore, most of the programming tools available for non-targeted analysis focus on environmental monitoring and not comparative samples, such as samples with and without treatment or exposure. An automated data analysis tool, which compares analyte concentrations before and after treatment or exposure to determine analyte formation or degradation, is also not currently available. Therefore, there is a need to automate GC×GC/ToF-MS data analysis of comparative samples.

In non-targeted analysis, a workflow to prioritize the identification and confirmation of candidate compounds is needed. Studies using GC×GC/ToF-MS analysis to screen for halogenated organic compounds in marine environments relied on halogenated isotopic clusters, fragmentation patterns indicating loss of halogen, and mass spectral library matches to identify, prioritize, and confirm candidate compounds [10,22,23]. The prioritization of chemicals can be based on several factors, including whether there is fragmentation data for known standards, electron impact (EI) mass spectral databases (such as NIST EI mass spectral library), previously published mass spectra data, and retention time match with known standards [11,23–25]. For EI, there are well developed strategies to predict compounds based on mass fragments, several fragments indicating unique functional groups, or chemical structures [26].

The objective of this study was to develop an automated method to streamline data analysis for GC×GC/ToF-MS non-targeted analysis of comparative samples by reducing the data analysis time. We developed a user-friendly Python^™ script (OCTpy) that automates GC×GC/ToF-MS data interpretation by refining the list of suggested analytes generated by the LECO^® ChromaTOF^® software. In a previous study, coal tar-contaminated soil samples were aerobically remediated in a laboratory-scale bioreactor and measured for toxicity pre- and post-bioremediation [27]. Although concentrations of parent-PAHs, which are environmental contaminants with human health implications,[28–30] decreased post-bioremediation, the soil toxicity increased, as evidenced by increase in genotoxicity from the DT40 chicken lymphocyte bioassay and observed morphological malformations from the zebrafish (Danio rerio) developmental toxicity testing.[27] We used OCTpy to determine analytes formed during bioremediation of soil samples with data generated by GC×GC/ToF-MS.[27] We benchmarked the OCTpy results with those obtained through human manual analysis to establish the reliability of OCTpy results and the effectiveness of OCTpy in reducing GC×GC/ToF-MS data analysis time.

2. Material and Methods

2.1 Sample Extraction and Analysis

It was hypothesized that the increase in soil toxicity was due to the formation of oxygenated transformation products of 3- and 4-ringed parent PAHs and other unknown analytes during bioremediation.[27] To test this hypothesis, extracts of pre- and post-bioremediation soil samples were analyzed using GC×GC/ToF-MS. Briefly, soil samples were extracted in hexane:acetone (75:25 v/v) using pressurized liquid extraction and fractionated into six fractions (A–F) based on polarity.[27] Laboratory blanks (consisting of sodium sulfate) also went through the same extraction procedures. To analyze analytes of interest that were not GC amenable, such as hydroxy-PAHs, aliquots of soil fractions were derivatized with N-tert-Butyldimethylsilyl-N-methyltrifluoroacetamide (MTBSTFA).[31] Both underivatized and derivatized fractions were analyzed in triplicate. Further details of the analytical procedures are available elsewhere.[11,27]

Pre- and post-bioremediation soil fraction extracts, along with blanks, were analyzed using an Agilent 7890 GC (Palo Alto, CA) coupled to a LECO^® Pegasus^® 4D time-of-flight mass spectrometer (St. Joseph, MI) with Restek Rtx-5 (35 m × 0.25 mm I.D. × 0.1 μm film thickness with 5-m integrated guard column) and Rxi-17 (1 m × 0.1 mm I.D. × 0.1 μm film thickness) (Bellefonte, PA) as the first and second dimension columns, respectively.[11,32,33] The injection volume for each sample was 2 μL, and the inlet temperature was held at 300 °C. For the first dimension, the oven was held at 60 °C for 1 min, ramped at 6 °C/min to 300 °C, held for 3 min, ramped at 20 °C/min to a final temperature of 320 °C, and held for 15 min. For the second dimension, the oven was first held at 85 °C for 1 min, ramped at 6 °C/min to 320 °C, held for 3 min, ramped at 20 °C/min to 340 °C, and held for 15 min. The non-moving quad-jet dual-stage modulator temperature was 35 °C higher than the temperature in the first-dimension column, with a modulation period of 3.5 s. Transfer line and ion source were kept at 285 °C and 250 °C, respectively.

GC×GC/ToF-MS data for the soil bioremediation extracts were analyzed using LECO^® ChromaTOF^® software (St. Joseph, MI)[7,14,34] version 4.50. LECO^® ChromaTOF^® software has peak alignment, baseline correction, and peak deconvolution capability, but the program does not automate peak area difference calculations in chromatograms from comparative samples, resulting in additional data analysis time. Matches between the second dimension peaks and peaks from individual replicates were set at 500 (50%).[35] The minimum signal-to-noise (S/N) ratio for peak finding was 50. The LECO^® ChromaTOF^® software’s add-in feature, Statistical Compare (SC) (St. Joseph, MI), [34,36,37] was used in the soil bioremediation study. The LECO^® ChromaTOF^® software returned a list of suggested analytes based on mass spectral similarity matches to the National Institute of Standard and Technology (NIST) 2011 library. Based on the SC’s feature that identifies statistical differences in groups, we compared peaks assigned to three groups: blank, pre-, and post-bioremediation soil extracts. The output files, containing potential analytes of interest, were exported as text files (*.txt). Henceforth in this manuscript, “suggested analytes” refers to the list of analytes from the LECO^® ChromaTOF^® software’s SC output based on MS NIST 2011 library matches, while “candidate analytes” refers to the list of analytes from manual analysis or as an OCTpy output.

3. Theory

3.1 Manual Data Analysis

Figure 1 shows the workflow for manual data analysis. If a peak in the post-bioremediation extract increased by 1.5 fold, the analyte was interpreted as a candidate analyte formed during bioremediation and could potentially be responsible for the observed increase in toxicity post-bioremediation. These peaks were visually inspected, and only those peaks that were chromatographically resolved, and without severe tailing, were selected as candidate peaks. This manual workflow was repeated for each analyte, in each derivatized and underivatized fraction, and took many hours to accomplish.

Figure 1. Manual Analysis Workflow.

The manual workflow used to select candidate analytes from a given GC×GC/ToF-MS data output analysis from the LECO^® ChromaTOF^® software’s Statistical Compare feature. This process can take several hours per sample, depending on the complexity of the sample.

3.2 Automated Data Analysis

Figure 2 shows the workflow for automated data analysis with OCTpy, which relies on SC and acquired data input files from LECO^® ChromaTOF^® software. The SC input files contain data on the average peak area, for every suggested analyte, in every treatment group. For the purpose of this script, it is necessary to export the following four properties into a text file for every sample: (1) Analyte name (“Name”), (2) Quantifying mass-to-charge (m/z) ion (“Mass”), (3) Average quantifying peak area (“Area Average”), and (4) Average retention time (“R.T. Average (s)”). Having the area and RT average parameters in the datasets are advantageous because different experimental designs can have different numbers of replicates. There were three SC input files for every fraction in the soil bioremediation study (i.e., blank, pre-, and post-bioremediation files).

Figure 2. Automated Analysis Workflow.

The automated workflow used to select candidate analytes from a given GC×GC/ToF-MS data output analysis from the LECO^® ChromaTOF^® software’s Statistical Compare feature. This process only take several minutes per sample, depending on the complexity of the sample.

OCTpy also requires an acquired sample input file prior to the SC feature analysis from one replicate, in one of the treatment groups, in order for the script to run. Serving as a quality control step, the acquired file is a reference file that verifies the presence and precision of the quantifying m/z ion and second dimension RT for each suggested analyte. The second dimension RT is used because there can be variations in the first dimension RT that arises from modulation time differences if different chromatogram slice is used as the base. The following three properties are pertinent for the script: (1) Molecular weight of the suggested analyte (“Exact Mass”), (2) First and second dimension retention times (“R.T. (s)”), and (3) Quantifying MS ion (“Quant Masses”). We used a replicate from the post-bioremediation sample for the acquired sample input file. Instructions for obtaining the SC input and the acquired sample input files for OCTpy are available in Appendix A.

4. Results and Discussions

OCTpy automates peak area comparisons of suggested analytes between comparative samples from the SC feature in LECO^® ChromaTOF^® software, thereby facilitating analyte selection by the analyst and shortening the data analysis times. Depending on the number of candidate analytes identified by the SC feature, the script returned results within seconds (Intel^® Core^™2 Quad CPU, 8 GB RAM, Windows^® 7, 64-bit), whereas manual data analysis required several hours for the same candidate analytes. Although preparing input files for OCTpy can take several minutes, analysts would still spend less time analyzing the peak area data by using OCTpy, compared to manual analysis. This is because OCTpy filters out suggested analytes that do not satisfy the desired criteria, such as those for which peak areas decreased following soil bioremediation. Following peak area comparison in a manual analysis, the analyst must meet additional criteria for identification by comparing the peak shapes of candidate analytes (Figure 1). However, we found that peak shape did not alter the results of OCTpy outputs and peak shape is not correlated with the toxicity of an analyte. Therefore, peak shape was excluded from the OCTpy script. The choice to either further review the peak shapes of candidate analytes from OCTpy with LECO^® ChromaTOF^®, or to use the output from OCTpy as the final result, is left to the discretion of the analyst. Once peak areas are compared, OCTpy parses through the acquired sample input file and checks each candidate analyte for the presence of the quantifying m/z ion or second dimension RT to the nearest tenth of a second (with an adjusted range of ±0.1 s to accommodate RT shift). For instance, if a candidate analyte has RTs of 100.789 s and 2.34 s, and the reference dataset contains RTs of 100.789 s and 2.429 s, then the candidate analyte will be included in the output. We successfully tested OCTpy on computers with Microsoft Windows^®, Mac^®, and Linux^® operating systems. The OCTpy source code and instructions for running OCTpy from the Windows^® command prompt are available in Supplementary materials section.

OCTpy outputs are available in the SI. We contrasted the candidate analytes identified using OCTpy to those identified by manual analysis, which included peak area comparison and individual peak review. Only analytes with chemical names (i.e., minimum similarity match with 2011 NIST library before name assigned > 700 (70%)) were selected for further analysis. There were a total of eight paired datasets, each consisting of a blank, pre-, and post-bioremediation samples from four soil extract fractions. Only four soil extract fractions were analyzed using GC×GC/ToF-MS, because increases in toxicity post-bioremediation were only measured in these four fractions.[11,27] Four of the eight datasets were from derivatized extracts, while four were from underivatized extracts. Among the eight paired datasets, the number of suggested analytes from the SC feature ranged from 170 (Fraction D) to 1,650 (Fraction C Derivatized) (Table 1). From the initial SC lists, OCTpy selected 20 (Fraction D) to 307 (Fraction D Derivatized) analytes having increased peak areas following bioremediation. As mentioned, once a list of candidate analytes is generated by OCTpy, the analyst is left with the option to either further examine the peak shape of each suggested analyte with the LECO^® ChromaTOF^® software or to proceed with the existing data and to look for specific groups of analytes, such as PAH transformation products.

Table 1. Manuals vs. Automated Result Comparisons.

Comparisons of candidate analytes selected by either manual analysis or OCTpy from the the LECO^® ChromaTOF^® software’s Statistical Compare (SC) feature list. Percent match between manual analysis and OCTpy was calculated based on the percent of the ratio between the number of candidate analytes that were selected in both OCTpy and manual analysis. For example, in fraction D, although OCTpy selected 20 candidate analytes, the 7 candidate analytes selected from manual analysis were also included in the OCTpy candidate analytes list, which resulted in 100% match.

Remediated Soil Fractions	Initial Number of Suggested Analytes from SC	Number of Candidate Analytes Identified by Manual Analysis	Number of Candidate Analytes Identified by OCTpy	Percent Match between Manual Analysis and OCTpy (%)
C	205	23	38	70
C-Derivatized	1,650	24	277	63
D	170	7	20	100
D-Derivatized	1,240	27	307	63
E	1,060	10	195	90
E-Derivatized	1,010	12	168	83
F	750	6	93	83
F-Derivatized	560	7	151	100

In contrast, manual analysis identified 6 (Fraction F) to 27 (Fraction D Derivatized) candidate analytes with increased peak areas following bioremediation. The manual lists of candidate analytes were shorter than the OCTpy list of candidate analytes because they underwent manual peak shape review with the LECO^® ChromaTOF^® software and OCTpy did not. For all eight data sets, OCTpy results contained 63–100% of the final list of candidate analytes (n = 87) generated by manual data analysis (Table 1). OCTpy generated more candidate analytes than manual data analysis of the initial list of suggested analytes from the SC feature, because peak shape was not included in the analyte selection criteria by OCTpy. OCTpy was able to select, on average, 18% of candidate analytes from the initial SC list output in a matter of several minutes. Even if an OCTpy user chooses to examine the peak shapes of the candidate analytes, the number of analytes to be examined is significantly reduced by executing OCTpy first, which results in decreased data analysis time. For example, the number of suggested analytes, from the SC feature analysis, in fraction D-Derivatized were reduced four fold once OCTpy was executed, which allowed the analyst to start the analysis with much fewer suggested analytes.

Twenty nine candidate analytes from manual analysis of the eight soil bioremediation data sets were not found in the OCTpy outputs. Several factors contributed to this. First, no peak areas (no value) in the chromatograms were replaced with the value “0” manually, whereas in the LECO^® ChromaTOF^® software SC input files, used in the OCTpy script, the fields were left blank. For example, for a dataset in which a given treatment group has three peaks with area values and one peak without, the analyst would manually calculate the average with n = 4, whereas LECO^® ChromaTOF^® and, consequently, OCTpy, would calculate the average with n = 3. This difference in how the average peak area was calculated explained why the seven candidate analytes that were selected manually in Fraction C were omitted by OCTpy. Given that different analysts treat missing values differently, it is imperative to create a standard method to analyze data that contain peak areas with no values. We recommend to using empty fields instead of inserting “0”. Since LECO^® ChromaTOF^® software calculates average peak areas without replacing any with zero, we propose OCTpy as the method of choice when analyzing LECO^® ChromaTOF^® software data, because it standardizes the calculation of average peak area.

Following close examination of the raw data, OCTpy excluded the remaining twenty two candidate analytes identified by the manual data analysis for one of the two following reasons (both of which were largely attributed to human error during data analysis): 1) the peak areas in the blank fractions were higher than those in the pre-bioremediation fractions, or 2) the peak areas in the pre-bioremediation fractions were higher than the post-bioremediation fractions (Appendix B). Detection of these discrepancies demonstrated OCTpy’s ability to account for human errors prior to qualitative identification of candidate analyte structures. However, future studies that use OCTpy should consider an assessment where manual and automated methods are compared, using a number of authentic standards with two different known concentrations, to reduce the potential for discrepancy between the two methods.

OCTpy’s capacity to reduce the number of candidate analytes within several minutes, along with its ability to reduce the potential for human error, represents a significant improvement over manual analysis that can take several hours. OCTpy gives the user the flexibility to not only identify analytes with increased peak areas following treatment in comparative samples, but also those with decreased peak areas. In addition, the user can also choose to either analyze differences between two groups, such as pre- and post-bioremediation, or three groups, such as blank, pre-, and post-bioremediation.

Once candidate compounds are identified, the analyst must confirm the structure of compounds using pure standards [24,38,39]. Using standards, Chibwe et al. identified N-(5-amino-4-cyano-1-pyrazolyl)phthalimide and 4-methylphthalic anhydride as the two compounds in fraction F with increased peak area post-bioremediation based on GC×GC/ToF-MS analysis.

5. Conclusions

OCTpy is a python script designed to automate analysis of GC×GC/ToF-MS data obtained from LECO^® ChromaTOF^® software. OCTpy uses SC feature input files that were generated by the LECO^® ChromaTOF^® software. Automated data analysis can now be completed within several minutes using OCTpy, making OCTpy an efficient platform that can be adopted by analysts who are using GC×GC/ToF-MS and the SC feature from the LECO^® ChromaTOF^® software. Because the list of candidate analytes are curated using the SC feature, only the most statistically significant analytes are included. However, the analyst should ensure that the MS fragmentation pattern of the candidate analyte is the same as the NIST EI mass spectral library.

While the decision to review peak shape following OCTpy analysis is left to the analysts, OCTpy provides assistance by making the analysis of large datasets practical and less time-consuming [41]. In cases when there are a significant number of candidate compounds that need to be confirmed for their toxicity, high throughput screening, such as the zebrafish developmental toxicity testing [42,43], or in silico search using the U.S. Environmental Protection Agency’s Tox 21 approach [44,45], are useful to reduce testing time. OCTpy also reduces human error, which is considered the most common source of error in quantitative analysis.[46] OCTpy’s capabilities are also unique, because the script aids researchers in data analysis and screening of candidate analytes with increased or decreased peak areas for comparative samples. In the spirit of open-source,[47] OCTpy is available on request, allowing further developments by users who are interested in examining the source code. Integrating OCTpy into the overall workflow of GC×GC/ToF-MS for non-targeted analysis[24] can also provide a standard method, across laboratories and research groups, for data analysis.

Supplementary Material

Appendix

Appendix A. Information on how to obtain input files for OCTpy and to run OCTpy

Appendix B. OCTpy soil bioremediation study results