A FAME mixture consisting of a set of 22 fatty acid methyl esters of chain lengths from C8–C30 was purchased in the form of the Fiehn GC/MS Metabolomics Standards Kit (7 ampoules: 1 × 0.5 mL FAME/d27 mixture, 1 × 0.5 mL pyridine, 1 × 0.5 mL MSTFA/1 % TMCS and 4 × 1.2 mL d27 mystric acids mix) from Agilent Technologies (Santa Clara, CA, USA). O-methylhydroxylamine hydrochloride, MSTFA (N-methyl-N-trifluoroacetamide) with 1% TMCS (trimethylchlorosilane), TSP (trimethylsilylpropionic acid-d4, sodium salt) and pyridine (silylation grade) were purchased from Sigma Aldrich (St Louis, MO, USA).

Plants samples were collected in several Ecological stations in São Paulo State, Brazil as part of the storage of 2,000 plant extracts by Brazilian Biodiversity Virtual Institute Program (www.biota.org.br). Giselda Durigan identified all species and vouchered specimens were deposited at the São Paulo State Botanical Institute herbarium (SP) as shown in Table 1.

The plants were separated into leaves and stems, dried at room temperature and ground in a Wiley mill. The extraction procedure was chosen on the basis of similar conditions previously reported (Ju and Howard, 2003; Bergeron et al., 2005). Extractions were conducted using a Dionex ASE 100 system (Oakville, ON, Canada) with stainless steel vessels (66 mL) using 0.5 g of dry ground plant, and ~60 mL EtOH at 60°C and 1500 psi for 15 min. The extracts were dried using a vacuum evaporator (Eppendorf, Hauppauge, NY, USA).

All samples underwent a two-step derivatization procedure before GC-MS analysis (Gu et al., 2013). Initially, methoximation was performed to protect aldehydes and ketones and to inhibit the ring formation of reducing sugars. O-methylhydroxylamine hydrochloride solution (10 μL), prepared using 40 mg mL⁻¹ O-methylhydroxylamine hydrochloride, (Sigma-Aldrich no. 226904—98.0%) in pyridine (99.9%) was added to the samples, and the mixtures were kept at 30°C for 90 min. Next, 90 μL N-methyl-N-trimethylsilyltrifluoroacetamine with 1% chlorotrimethylsilane (MSTFA+1% TMCS) was added to the samples and kept at 37°C for 30 min to allow the trimethylsilylation of acidic protons. Subsequently, 2.0 μL of the FAME mixture was added to each sample to provide retention time indices. The solutions were vortexed and transferred to GC-MS glass vials for analysis (Kind et al., 2009; Gu et al., 2013).

Experiments were performed on an Agilent 7890A GC-5975C MSD system (Agilent Technologies, Santa Clara, CA) using a DB5-MS+10m Duraguard Capillary Column (30 m × 250 μm × 0.25 μm) as the stationary phase. The GC parameters used were as follows: split injection (1.0 μL sample at 100.0°C, 1.0 min—split ratio of 10:1); He carrier gas (40 cm s⁻¹ at constant velocity); 275.0°C transfer line temperature; oven temperature program: 1.0 min at 100°C, increased 20.0°C min⁻¹ to 200.0°C, then increased 3.0°C min⁻¹ to 325.0°C and held for 10.0 min, MS parameters: electron impact ionization at 70 eV, filament source temperature of 230.0°C, quadrupole temperature of 150.0°C, m/z scan range 50–600 at 2 spectra s⁻¹. Mass spectral signals were recorded after a 6.10 min solvent delay to avoid derivatization interferents, and turned off between 10.0 and 13.0 min to avoid saturation of the detector due to the high content of monosaccharides. A blank sample with the FAME standard mixture (FAME std) was also injected under the same GC conditions.

The analysis of the full-scan data files acquired by GC-MS was performed using the empirical method developed by Dromey et al. (1976) and employed by Stein (1999) using the AMDIS software package. The overall deconvolution process in AMDIS consists of three sequential steps: (1) noise analysis, (2) component perception and (3) model shape determination and spectrum deconvolution (Stein, 1999). The first step extracts the noise from GC-MS data file by empirically calculating a noise factor using the median abundance level of a representative background segment obtained from 13 scans.

Component perception (2), identifies individual chromatographic components by considering ions that have maximal intensities at the same time (i.e., the same or similar scan). It is achieved through a process which sequentially examines individual peak maxima using a pre-set number of scans (4–32) in the forward and reverse directions.

Step (3) determines the peak shape for each perceived component and performs the extraction of “pure” MS spectra. A peak shape model is determined by considering the sum of individual ion chromatograms that maximize together and that have sharpness values within 75% of the maximum value for a particular component, as defined by: (1)sharpness=(Amax-An)(n×NfAmax) where A_max is the maximum abundance, A_n is the abundance from a pre-set number of scans n and N_f is noise factor. Finally, the spectrum is recovered using a least-squares method where each m/z value is individually fit to the model profile, allowing a linear baseline: (2)A(n)=a+b×n+c×M(n)+d×Y(n) where A_(n) is the abundance during scan n. a, b, and c are derived constants, and M_(n) is the abundance of the model profile during scan n. Box 1 illustrates the metabolic identification process using AMDIS.

Ratio analysis is designed to work using single datasets that contain multiple MS spectra for the same metabolite (Gu et al., 2013). For peaks that originate from the same compound, under the same experimental conditions, their MS peak intensity ratios across the chromatographic peak should be relatively constant. In addition, the standard deviations of those ratios should be small (zero in principle; Gu et al., 2013).

The procedure for calculating the RAMSY spectrum was described previously (Gu et al., 2013). Briefly, one isolated peak in the mass spectrum is selected as the driving peak. The intensity value of this selected driving peak is divided by the intensities of all the other peaks in the spectra one at a time, as shown in Equation (3): (3)Di,j=Xi,jXi,k where, the vector X_i is the ith spectrum of a set of n MS spectra, and the jth data point of m total points in that spectrum is denoted as X_{i, j} (X_{i, k} is the driving peak). D is the ratio matrix of dimension n × m.

The RAMSY values, denoted as an m-element vector R, are the quotients of means and standard deviations across columns of D. The standard deviation is zero for the driving peak itself; therefore, its RAMSY value is pre-defined (e.g., the value of the highest RAMSY ratio). The other RAMSY values are calculated as elements of the vector R as follows: (4)Rj=1n∑i=1nDi,j1n∑i=1n(Di,j-1n∑i=1nDi,j)2

Since a ratio's standard deviation is used as the denominator, a small standard deviation will produce a large reciprocal value, generating a peak (in principle an MS peak from the same compound as that for the driving peak). In general, the MS peaks from interfering compounds will generate large standard deviations and thus small RAMSY numbers, similar to noise values. Notably, RAMSY values are dimensionless.

GC-MS metabolite identification was performed considering two independent parameters: the Linear Retention Index (LRI) and the mass spectrum (MS) similarity profile. LRIs were based on linear regression using FAME internal standards retention times according to Van den Dool and Kratz, and are shown in Figure 1 (Van Den Dool and Kratz, 1963).

The MS similarity profiles were calculated by comparison of AMDIS and RAMSY deconvoluted spectra using two MS databases, FiehnLib (University of California, Davis, CA, USA, Agilent webpage) and the Golm Metabolome database (Max Planck Institute, Potsdam–Golm, Germany). The similarities between samples and databases were calculated using the same algorithm as those reported for the NIST library (Stein and Scott, 1994). Briefly, we first obtained the “angle” between the two spectra:

M is the m/z value, and A_S and A_U are the base-peak normalized abundances of the peaks in the standard spectrum and unknown spectrum, respectively. Next, F₂ is calculated:

F₂ is based on relative intensities of pairs of adjacent peaks present in both spectra. N_U&S is the number of peaks common to the unknown and standard spectra, and n = 1 (−1) if the first abundance ratio is less (larger) than the second. The Match Factor (MF) is then calculated as follows: (7)MF=1000NU+NU&S(NUF1+NU&SF2)

A perfect match results in an MF-value of 1000; spectra with no peaks in common result in a value of 0.

Previous studies on automated metabolite identification efficacy using AMDIS showed relatively high (27.8–32.8% false positive rates (on average) using MF 700–900 (Aggio et al., 2014). For that reason, we considered a positive identification only with LRI errors ≤ 5% and MS similarity profiles with MF ≥ 700. The RAMSY algorithm was applied for chromatographic regions with detected metabolite MFs in the range of 700–790. Box 2 illustrates the metabolic identification process using AMDIS-RAMSY.

The use of computational methods can assist the identification of known metabolites by extracting the signals from co-eluted GC-MS components (Du and Zeisel, 2013). AMDIS is a freely available software package that applies arbitrary rules for peak deconvolution and performs identification in an integrated matching system combined with the NIST standard reference and others databases (Stein, 1999). The designed dereplication protocol started with the application of a 2v⁵⁻¹ fractional design on AMDIS parameters: component width (8 or 32), number of adjacent peaks (0 or 2), resolution (low or high), sensitivity (low or high) and shape requirements (low or high). The results were evaluated according to the compound detection factor (CDF) calculated by the proposed heuristic Equation (8), (8)CDF=A×(AB)x

CDF provides the optimized ratio between the number of detected (A) and identified compounds (B) by reducing the negative effects of variable over-fitting due to the inclusion of noise and/or false components. “A” represents the identification power derived from the library dataset extension, while the “(A/B)^x” ratio expresses a penalty to conditions where a large number of peaks are detected but not identified. The “x” value depends on how important the constraint factor is to the model (typically x = 3). The best results occur when the relation between “A” and “B” is simultaneously increased. On the other hand, when only “B” is increased, CDF is reduced. The optimized deconvolution and identification parameters by AMDIS using the CDF are shown in Table 2.

The CDF indicated that high “adjacent peak subtraction,” low “component width,” and low “sensitivity” generated the best deconvoluted chromatograms, regardless of the sample, whereas “resolution” and “shape requirements” showed particular response based on the metabolic composition.

Optimized AMDIS deconvolution yielded ~200 components per sample, of which ~100 were putatively identified based on mass spectral MF and LRI correlations, as shown in Table 3. In general, the components that could be identified included amino acids, organic acids, fatty acids, carbohydrates, sugar alcohols, phytosterols, and specialized metabolites such as flavonoids, triterpenes, phenolics, polyamines, and related compounds.

After performing the optimization of AMDIS-based GC-MS dereplication, we applied RAMSY in regions that showed low MFs due to high peak overlap, low intensities and/or noise effects. In the first application of RAMSY, we selected a broad peak between 7.45 and 7.60 min for the Couepia grandiflora plant species sample (Chrysobalanceae). According to AMDIS, this peak represented the overlap of pyroglutamic acid (Rt 7.48 min, MF = 750), dodecanoic acid methyl ester from the FAME standard mixture (Rt 7.49 min, MF = 810) and threonic acid (Rt 7.53 min, MF = 710). We also observed the presence of unidentified low-intense ions 304 m/z and 174 m/z. We selected the 15 scans that formed the GC peak at 7.45–7.60 min for RAMSY analysis. The RAMSY spectrum calculation was performed using the fragment ion 156 m/z as the driving peak for pyroglutamic acid, 87 m/z for dodecanoic acid methyl ester, 292 m/z for threonic acid and 304 m/z for the unidentified compound. The averaged EI-MS spectrum was filtered with the RAMSY values (only those MS peaks with top RAMSY values were shown) and compared with MS libraries considering MF. RAMSY correctly identified ions from the three metabolites identified by AMDIS, as depicted in Figure 2. The application of RAMSY using 304 m/z as the driving peak provided a new mass spectrum suggested as 4-aminobutyric acid, based on the NIST MS database (MF = 840). These results provided evidence for the capabilities of RAMSY to recover low-intense ions in co-elution. RAMSY also provided MF values > 900 for the other compounds detected by AMDIS—dodecanoic acid methyl ester (MF = 920), pyroglutamic acid (MF = 970) and threonic acid (MF = 920), attesting to the ability of RAMSY to provide deconvolution enhancement.

In another application of RAMSY, the chromatographic peak between 37.75 and 38.00 min observed in Solanum americanum led to the dereplication of three overlapped compounds by AMDIS, Figure 3. The molecules were identified as α-tocopherol (MF = 500), octacosanoic acid methyl ester (MF = 770) and raffinose (MF = 600). The RAMSY spectrum was calculated using the 26 scans that compose the GC-MS peak, by selecting driving peaks at 237 m/z, 87 m/z and 361 m/z. RAMSY improved the deconvolution, resulting in increased MF values for each of the metabolites: α-tocopherol (MF = 900), octacosanoic acid (MF = 870) and raffinose (MF = 900).

The comparison between RAMSY and AMDIS using the 2v⁵⁻¹ optimization design experiments is shown in Figure 4 and indicates that the use of AMDIS alone consistently generated lower MF values and more false negatives. The application of RAMSY in combination with AMDIS assisted by CDF improved the identification of known metabolites in very complex biological samples.

In order to evaluate the identification efficiency, we compared the GC-MS raw data and processed metabolite profiles (AMDIS-RAMSY) by HCA, as depicted in Figure 5. It is possible to observe a taxonomic correspondence based on the metabolic content from the GC-MS raw data and the dataset with the identified metabolites after AMDIS-RAMSY (colored spots). The AMDIS-RAMSY results presented better taxonomical grouping compared to raw data due to the spectral metabolic deconvolution and suppression of noise and interferent effects. According to the AMDIS-RAMSY taxonomical grouping, amino acids and fatty acids were responsible for distinguishing between Solanaceae species, while for the Chrysobalanaceae species the differentiation was based mainly on carbohydrates and flavonoids, as evidenced in Figure 5.

DR is an officer at Matrix-Bio, Inc. The other authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.