Improved Segmented-Scan Spectral Stitching for Stable Isotope Resolved Metabolomics by Ultra-High-Resolution Fourier Transform Mass Spectrometry

Abstract

We have implemented a linear ion trap (LIT)-based SIM-stitching method for ultra-high-resolution Fourier transform mass spectrometry (FTMS) that increases the S/N over a wide m/z range compared to non-segmented wide full-scan (WFS) spectra. Here we described an improved segmented spectral scan stitching method that was based on quadrupole mass filter (QMF)-SIM, overcame previous limitations of ion signal loss in LIT. This allowed for accurate representation of isotopologue distributions, both at natural abundance and in stable isotope-resolved metabolomics (SIRM)-based experiments. We also introduced a new spectral binning method that provided more precise and resolution-independent bins for irreversibly noise-suppressed FTMS spectra. We demonstrated a substantial improvement in S/N and sensitivity (typically > 10-fold) for ¹³C labeled lipid extracts of human macrophages grown as three-dimensional (3D) cell culture, with detection of an increased number of ¹³C isotopologue ions. The method also enabled analysis of extracts from very limited biological samples.

Graphical Abstract

1. Introduction

Direct infusion nanoelectrospray ultra-high-resolution Fourier transform mass spectrometry (DI nESI UHR-FTMS) instruments enable analyses of crude metabolite extracts without prior separation while achieving the high mass accuracy and resolution needed for determining of molecular formula in complex mixtures. However, high-resolution ion trap instruments such as ion-cyclotron resonance (ICR) and Orbitrap™ MS can experience loss of spectral quality due to space charge effects. These effects decrease the dynamic range, which compromises wide scan spectral quality. This is particularly problematic for biochemical pathway-tracing approaches such as stable isotope-resolved metabolomics (SIRM) [1–3], which can result in 40 or more metabolite isotopologues within a range of 40 m/z. Some of these isotopologue ions are not detectable due to low S/N, and can heavily overlap with other constituents, thus requiring high S/N, accuracy, resolution wide scan MS1 acquisitions for proper analysis of isotopologue distribution (Figure S1) and metabolic interpretation [2, 4].

To address these issues, Southam et al. [5] developed the SIM-stitching method using ultrahigh resolution ICR-FTMS with a front end linear ion trap or LIT (Thermo LTQ-FT). Their approach took advantage of a well-recognized property of MS that higher S/N can be achieved using a narrower m/z range for a fixed number of ions. Their method consists of multiple narrow SIM segments, each with a set, optimum automatic gain control (AGC) target value, which controls the number of ions injected into detectors. These SIM segments are then stitched together post-acquisition to produce a wide range spectrum, which is necessary for multivariate statistical analysis of metabolomic data and useful for visual inspection [5]. Generating single, contiguous spectra is equally important when searching for interspersed isotopologue peaks from different metabolites (Figure S1c) [6]. The spacings for these isotopologue ions can span 40 m/z or more, with isotopologue clusters dispersed across the entire spectral range (Figure S1a). We have implemented Southam’s method for our DI nESI UHR-FTMS analyses of SIRM lipid data, where the spectra had very high peak density and dynamic range (Figure S1). Implementation of this method on our QMF-SIM-based instrument helps eliminate the edge effect [5, 7]. This is because ion isolation utilizing QMF-SIM does not suffer from the space charge effect seen in the LIT, which contributes to the edge effect. In addition to the edge effect, we encountered another limitation of the method, i.e. applicability of a conventional noise filtering method only for full-profile spectra (true raw spectra with intact baseline noise), which are unavailable in some commercial form of Orbitrap instruments [8].

Many QMF-FTMS instruments automatically truncate the baseline noise and provide only the reduced-profile spectra, which lack the critical information for the conventional noise reduction process [5, 8]. For the reduced-profile spectra, a recently reported intensity-independent noise filtering method [9] can be substituted for the conventional method since it does not require baseline noise information from full-profile spectra. While this method provides a reasonable noise threshold with minimal introduction and removal of false signals, the intensities are unfortunately distorted due to the introduction of noise signals when averaging spectra followed by a resolution-dependent spectral binning [9]. We employed a new spectral binning method that accounts for and leverages the way computers store decimal numbers to overcome the intensity distortion introduced by the resolution-dependent method.

Here we describe an improved segmented-scan spectral stitching method for QMF-FTMS instruments that provide only reduced-profile spectra and demonstrate it on an Orbitrap Fusion™ Tribrid™ (Thermo Scientific). In addition, the method can be optimized based on specific sample ion distribution profiles. We demonstrated this sample-specific silhouette segmented-scan spectral stitching (S⁷) on the extremely dense, highly demanding lipid analysis for SIRM studies, including human macrophages grown as 3D cultures.

2. Materials and Methods

2.1. Materials

2.1.1. Isolation of human monocytes

Human whole blood (WB) was obtained from healthy volunteers as part of a lung cancer metabolism study according to a University of Kentucky IRB approved protocol (IRB# 43807). Monocytes were isolated from WB using the RosetteSep™ human monocyte enrichment cocktail kit (StemCell Technologies Inc., Cambridge). Each mL of WB was mixed and incubated for 20 min with 2 μL of 0.5 mM EDTA, and 50 μL of the RosetteSep™ monocyte negative selection cocktail. The WB was diluted 1:1 with phosphate-buffered saline (PBS) (Ca²⁺ and Mg²⁺ free) + 1× antibiotic-antimycotic (Thermo fisher) + 2 % endotoxin-free fetal bovine serum (FBS) and then added to a Ficoll-Hypaque (GE) gradient in SepMate™-50 tubes (StemCell Technologies Inc.), followed by centrifugation at 1,200 ×g for 20 min. The top layers were pooled and then spun for 10 min at 400 ×g. The resulting cell pellets were resuspended in the same 2 % FBS-PBS dilution buffer and washed twice via centrifugation at 600 ×g and 120 ×g for 10 minutes each. Finally, the cells were counted and seeded into 6-well plates at a density of 3 – 5 × 10⁶ cells per well in Dulbecco’s Modified Eagle Medium (DMEM) supplemented with 0.2 % glucose, 2 mM glutamine, 10 % FBS, 1× antibiotic-antimycotic, and 50 ng/mL of monocyte colony stimulating factor (M-CSF) (M0 medium).

2.1.2. Macrophage spheroid formation and ¹³C₆-Glc SIRM experiment

Isolated human monocytes were differentiated into macrophages in 6-well plates for 5 days in M-CSF supplemented medium at 37 °C-5% CO₂ before treatment with NanoShuttles™-PL (NS; n3D Biosciences) at 1 μL per 10⁴ cells overnight according to the vendor’s protocol. After 16 h, NS-loaded cells were then detached using accutase, counted, and bioprinted for 1 h in cell-repellent flat-bottom 96-well plates (Greiner Bio-One) at a density of 200,000 cells per well using the magnetic spheroids drive (n3D, Biosciences) for spheroid formation [10]. Cells were either not activated (M0) or classically activated to the M1 subtype by treating the cells with 100 ng/mL lipopolysaccharide (LPS) plus 20 ng/mL interferon gamma (IFNγ) for 3 days [11, 12]. For the SIRM experiment, cells were incubated for another 24 h in fresh M0 medium with glucose replaced by 0.2 % ¹³C₆-glucose.

2.1.3. Extraction of metabolites from macrophage spheroids

Macrophage spheroids were held with the 96-well magnetic holder and washed twice with cold PBS and once briefly (30 sec) with Nanopure water to remove salts. The spheroids were quenched in 200 μL of cold 70 % methanol per well, which also served to extract polar metabolites [10]. The remaining pellets were extracted for lipids by adding 600 μL of methyl tert- butyl ether (MTBE) and 180 μL of 100 % methanol to each pellet. The mixtures were vortexed at 3,000 rpm for 1 h at room temperature. After adding 210 μL of Nanopure water, the mixtures were vortexed for 30 sec and centrifuged at 14,800 ×g for 10 min at 4 °C. The top lipid layer was collected, and the residue was extracted again as described above. The two pooled lipid extracts were lyophilized and dissolved in 200 μL of chloroform : methanol : butylated hydroxytoluene (2:1:1 mM) solvent mixture before FTMS analyses.

2.2. Methods

2.2.1. AGC target optimization

An nESI spray solvent mixture of isopropanol : methanol : chloroform : ammonium formate (4 : 2 : 1 v/v : 20 mM) containing 2,800-fold diluted deuterated-lipid (D-lipid) standard mixture (SPLASH™ LipidoMix®, Avanti Polar Lipids, Alabaster, AL) was prepared. The mixture was injected into the FTMS (THERMO Scientific™ Orbitrap Fusion™ Tribrid™) via nESI (Advion TriVersa NanoMate®) in positive ion mode. The applied voltage was 1.5 kV with 0.4 psi head pressure and was controlled by ChipSoft 8.3.3 software (Advion). AGC targets of 1 × 10³, 2 × 10³, 5 × 10³, 1 × 10⁴, 2 × 10⁴, 5 × 10⁴, 1 × 10⁵, 2 × 10⁵, 5 × 10⁵, and 1 × 10⁶ were selected for three different 100 m/z segments of 480–580, 625–725, and 740–840, which is where the D-lipid standards for calibration ([18:1(d7) lysophosphatidylcholine (LPC) + H]⁺: m/z 529.39936, [18:1(d7) cholesteryl ester (CE) + NH₄]⁺ : m/z 675.67794, and [15:0–18:1(d7)−15:0 triacylglycerol (TG) + NH₄]⁺: m/z 829.79845) are centered. Ion injection times were 25, 50, 100, and 500 ms for AGC targets 1 × 10³–5 × 10³, 1 × 10⁴–5 × 10⁴, 1 × 10⁵–2 × 10⁵, and 5 × 10⁵–1 × 10⁶, respectively. All the spectra were acquired for 2 minutes with five microscans per scan at a measured R_{400 m/z} ≈ 370,000. The scans were averaged using Xcalibur 3.0.63 (THERMO Scientific™). Each segment was calibrated against the three D-lipid standards. The mass errors were obtained using the other D-lipid standards within each segment in combination with known sample contamination, such as polydimethylsiloxane (PDMS), polyethylene glycol (PEG), etc. [13, 14] (Table S1) using the in-house software Precalculated Exact Mass Isotopologue Search Engine (PREMISE) [4, 15].

2.2.2. Evaluating edge effects

The Thermo Fusion mass spectrometers include QMF, LIT, and Orbitrap; this enabled selection of QMF or LIT to directly compare ion isolation for SIM-scans. The nESI spray solvent mixture described above (2.1.1) was injected into the Fusion via the nESI in positive ion mode. Maximum ion injection time was 100 ms with an AGC target of 1 × 10⁵. All the spectra were acquired for 12 seconds with 5 microscans per scan, then the scans were averaged using Xcalibur. Intensity ratios of monoisotopic to m + 2 isotopologue ions of the D-lipids (18:1(d7) LPC (529.39936 / 531.40607), 18:1(d7) CE (675.67794 / 677.68465), and 18:1(d7)−15:0 TG (829.79845 / 831.80516)) at different locations in 100 m/z segments were measured in triplicate in a similar way to Weber et al. [7].

2.2.3. Mass spectrometry of human macrophage lipids from SIRM experiments

The non-polar phase extracted from the human macrophages was diluted 1 : 5 with the spray solvent mixture. The samples were injected into the FTMS via nESI in positive ion mode. The applied voltage was 1.5 kV with 0.4 psi head pressure and was controlled by the ChipSoft. Xcalibur was used to set up the S⁷ method. MS1 QMF-SIM was used to filter the segments to the desired m/z ranges. For the S⁷ method, each stored scan was composed of 1 microscan with a total acquisition time of 15 minutes including dummy scans. For WFS spectra, 5 microscans were accumulated per stored scan for a total acquisition time of 10.5 minutes including dummy scans. The AGC target was 1 × 10⁵ with 1000 ms of maximum ion injection time for both S⁷ and WFS. Fluoranthene (m/z 202.07771) was injected into the detector by the instrument to serve as a lock mass (EASY-IC™) for WFS spectra only.

2.2.4. Processing FTMS spectra

Xcalibur 3.0.63 was used for visual inspection of acquired spectral raw files and to extract an exact m/z list from the WFS spectra. MSFileReader 3.1 (Thermo Scientific™) was used to export a series of m/z lists from individual scans and the maximum ion injection time from both WFS and S⁷ spectra to Microsoft Excel spreadsheets. MSFileReader is a library with a component object model (COM) standard that supports read access to Thermo raw files and can be used in any programming language that supports the COM interface. Two-step internal calibration has been conducted for WFS spectra. The spectra were first automatically calibrated using fluoranthene by the instrument during the acquisition, and the acquired spectra were subsequently calibrated using a D-lipid standard [15:0–18:1(d7) PC + H]⁺ (m/z 753.61337) either by PREMISE or the Python script described below. Unique internal calibrants (Table 1) were used to calibrate each segment for S⁷ spectra. The instrument has been calibrated on a two-week basis using a calibration solution (Pierce™ LTQ ESI Positive/Negative Ion Calibration Solution). A Python script was used for centroided m/z list (what Thermo calls “label data”) extraction from .raw files, poor scan elimination and inter-scan normalization [5] using maximum injection time, alignment of m/z in an ascending order across scans within a segment, 32-bit floating point format-based spectral binning, m/z median and intensity average, m/z correction using internal calibrants, intensity-independent noise filtering [9], inter-segment normalization using intensity ratios of peaks mutually present in the overlapping regions, and stitching for S⁷ and WFS spectra from the macrophage samples (the last two steps are only for S⁷). The Python script is available at https://zenodo.org/badge/latestdoi/150272490. PREMISE was used for both WFS and S⁷ to assign observed m/z values to an in-house lipid database [4, 15] and common contaminants lists (Table S1) [14] for the mass error measurement.

Table 1. Parameters of the optimized S⁷ for lipid extracts of M0 and M1 macrophage spheroids cultured in ¹³C₆-glucose.

Number of microscan	Minimun number of scan	Duration time (min)	Low m/z	High m/z	Calibrant
					mass	Molecular ID	Molecular Formula		Adduct
1	4	0.09	150	350	223.06365	PDMS*	C6H18O3Si3		[+H]+
1	4	0.09	330	530	445.12002	PDMS	C12H36O6Si6		[+H]+
1	2	0.06	510	610	529.39936	18:1(d7) LPC	C26H52N1O7P1		[+H]+
1	2	0.06	600	700	675.67794	18:1(d7) CE 15:0-	C45H78O2		[+NH4]+
1	2	0.06	690	790	753.61337	18:1(d7) PC 15:0-	C41H80N1O8P1		[+H]+
1	1	0.03	780	830	829.79845	18:1(d7)-15:0 TG	C51H96O6		[+NH4]+
1	1	0.03	820	870	832.24053	PDMS	C22H66O11Si11		[+NH4]+
1	1	0.03	860	910	906.25932	PDMS	C24H72O12Si12		[+NH4]+
1	2	0.06	900	1000	980.27811	PDMS	C26H78O13Si13		[+NH4]+
1	4	0.09	980	1180	1054.29690	PDMS	C28H84O14Si14		[+NH4]+
1	4	0.09	1160	1360	1276.35327	PDMS	C34H102O17Si17		[+NH4]+
1	5	0.12	1340	1600	1424.39085	PDMS	C38H114O19Si19	[+NH4]+

^*PDMS: polydimethylsiloxane contaminant

2.2.5. Optimizing settings of sample-specific silhouette segmented-scans

We found five major factors to be considered when designing the S⁷ method. The first step was determining the best sizes of the segments. While Southam et al. [16] recommended using an optimized and fixed SIM segment width for the spectral stitching method, Taylor et al. [17] segmented a wide mass range into multiple different widths of SIM segments based on the sample-specific m/z density distribution. Specifically, they used SIM widths of 100 m/z for 600– 900 m/z regions with denser m/z population, and any segments outside the range were widened to avoid under fill of LIT and hence the ICR detector. Their data were similar to our experiences with lipid extracts from various types of cancer, including those from patients, where we observed a denser m/z population between 600 and 1,000 m/z [2, 3, 15, 18, 19]. The intensity log sum from the WFS spectra of macrophage lipid extracts was used to design segments with evenly distributed total ion population density (Figure 1a), which was used throughout this study. We used 12 segments with different widths in the mass range of 150–1,600 m/z (Figure 1a, b). In order to avoid any computational backlog and errors due to large numbers of scans (more than 150 scans per segment) being processed when running the Python script, the m/z range of 1,160–1,600 was separated into two segments despite its lower intensity log sum. The shorter segments were set to acquire proportionally lower numbers of scans in order to expedite coverage of the wide mass range [5] (duration time, Table 1). When setting the width of segments for SIRM experiments, the size of the overlapping regions should be considered to ensure the transfer of accurate isotopologue silhouette (Figure S1b, c). This also has the advantage of enabling more robust normalization between segments by providing many mutually present peaks within the region. The overlapping region can be set to less than or close to 10 m/z in width when natural abundance isotopologue profiling is expected, since the peak spacings of observable isotopologues in general do not span more than 5 m/z. Second, each segment must have internal standards for m/z calibration. Any standard mixture that is composed of repeating units of oligomers, such as the PEG standard [5], is potentially useful. It has been reported that peaks of common chemical background such as plasticizers have served as lock masses [20], thus any known contaminants [14] for a given sample workflow can be used as internal calibrants. We selected eight PDMS ions as internal calibrants in combination with ions from four D-lipid standards as unique external calibrants for the 12 segments as shown in Table 1.

Figure 1. Schematic of S⁷ optimized for the ion density silhouette of mammalian lipids.

(a) Average log intensity sum of 12 m/z ranges from WFS (113 scans averaged) spectra of macrophage M0 subtype lipid extracts (triplicate). Error bar is RSD. (b) Distribution of 12 segments of a single S⁷ cycle with different m/z widths and different number of scans overlap by 10 or 20 m/z. The mass range is m/z 150–1,600 with an AGC target of 1 × 10⁵. A total of 18 S⁷ cycles comprised a single analytical run.

Unstable spray may occur intermittently in DI nESI, across multiple scans. Although this unstable spray results in a loss of MS signal, we find the DI nESI is typically stable for about 20 minutes with 200–600 nL min⁻¹ of flow rate with up to 15 μL of injection volume, as evidenced by consistent spray current at around 100 nA for the types of samples in this study. Any major spray instability can introduce errors in normalization of intensities based on injection time [5] when more than 1 microscan per scan is used. For the Thermo MS instruments, including the Fusion, a “microscan” comprises a single cycle of ion injection and storage or scan-out of ions followed by detection, whereas a “scan” is the disc-stored spectrum from the average of specified microscans. Since the instrument sums the microscans at the transient level followed by Fourier transformation, not only is it impossible to know which one of the multiple microscans lost their signal during the nESI instability but also the ion injection times for the individual microscans are lost. Therefore, we set the number of microscans to 1 for every stored scan, which allows rejection of individual poor scans to more accurately normalize inter-scan intensities. The maximum ion injection time was set to 1,000 ms to accommodate wide-ranging differences in segments among samples. We observed that the only instances when injection time reached 1,000 ms occurred when the signal was lost due to the nESI instability. The Python script described here assumes scans with an injection time equaling the maximum injection time are poor scans that experienced an unstable nESI profile and ignores them during analysis at an early processing stage.

To adjust the number of scans of WFS to be proportional to the size of its m/z range (1450 m/z width), approximately 870 or more scans are required. However, with the current Python script, more than 150 scans in combination with wide segments (> 200 m/z width) can cause computational backlog and errors due to the extremely large number of peaks detected. Thus, acquiring a reduced number of scans by increasing the number of microscans (5 in this study) per scan was necessary. This is a compromise between number of scans and quality control of microscans based on ESI spray stability.

We iteratively rotated the acquisition of the segments, to minimize the effects of any DI nESI instability or sample composition stratification. We found that a total of 18 cycles of 12-segments (ca. 49 seconds per cycle), for a total of 15 minutes of acquisition time with the first 30 seconds allocated to the dummy scans (Table 1, Figure 1b) provided an acceptable compromise between speed and reliability. The dummy scans minimize the effects of unstable nESI in the first several seconds of spraying [7].

It is critical to use resolving power that is sufficient to resolve natural abundance isotopic profiles for compound identification [21, 22], i.e. R_{400 m/z} ≥ 200,000 for SIRM experiments [4, 23]. Although it has been reported that low mass accuracy combined with relative isotope abundance information produced better assignments than high mass accuracy alone [7, 24], this is true only if natural abundance isotopic distribution can be assumed; this is clearly not the case for SIRM (e.g. Figure S1). Thus, the full capacity of the instrument has been utilized throughout; R200 m/z = 500,000, corresponding to a measured R_{400 m/z} ≈ 370,000 that takes approximately 1 second per microscan.

Considering all the above factors, the parameters of the method can be tailored according to the sample silhouette of ion density distributions, or to preselected analytical targets as discussed in the Methods (2.2.5) and in Figure 1.

3. Results and Discussion

3.1. Optimization of AGC target for mass accuracy in Orbitrap Fusion

For acquiring extremely high density, wide-scan MS1 data, such as those obtained from the ¹³C lipid profiles of SIRM experiments, a high S/N is required to distinguish low abundance isotopologue peaks from noise while minimizing the mass error. In such SIRM experiments, it is crucial to keep the mass error to around 0.1 ppm to reduce the number of possible candidate molecular formulae. For example, at 0.1 ppm mass accuracy, the number of candidates is typically 1 when calculated for an m/z of 150 with the search pool limited to the elements CHONPS, whereas it increases to 10 and 32 candidates for the m/z values of 700 and 900, respectively. The candidate list increases approximately 10-fold at 1 ppm of mass accuracy [24]. In practice, a small improvement in mass accuracy can be achieved by setting lower AGC target values, at the cost of reduced sensitivity.

In our case, the mass error started to increase with increasing AGC targets from 2 × 10⁵ to 1 × 10⁶ while the error remained consistent with AGC targets of 1 × 10³ through 1 × 10⁵ (Figure S2). Thus, we chose 1 × 10⁵ as the AGC target value throughout this study to maximize the S/N while maintaining an acceptably low mass error (ca. 0.2 ppm). In practice, the AGC target can be tailored for the particular silhouette of ion density in each segment, to optimize the S/N for a given application.

3.2. Elimination of edge effects by using wide quadrupole mass filter (QMF) ion isolation width

Southam et al. reported the loss of ions at the ends of SIM segments using instruments with either LIT or QMF [5, 7, 16]. The mechanism behind the “edge effect” in LIT [25, 26] and QMF [27–29] has been well described.

As previously reported [5, 7], LIT-SIM suffered from considerable ion loss at both ends of the segments, while QMF-SIM did not (Figure 2). As seen in Figure 2, the intensity ratios of monoisotopic (m+0) to m+2 ions remained consistent even though the accurate masses of m+2 ions of CE (m/z 677.68465) and TG (m/z 831.80516) and the m+0 ion of LPC (m/z 529.39936) fell in the +/− 0.5 m/z ends of the QMF-SIM segments. This is presumably due to a more rectangular ion isolation efficiency profile of the extended QMF-SIM segment compared with the narrower ones (Figure S3) [27, 28]. With our Orbitrap Fusion instruments, setting a segment window width of at least 50 m/z minimized the ion loss when using QMF-SIM (data not shown).

Figure 2. Differences in edge effect between LIT and QMS based acquisitions.

Values in the boxes are ion intensity ratios of m+0 : m+2 of D-lipids (a) 18:1(d7) LPC, (b) 18:1(d7) CE, and (c) 18:1(d7)−15:0 TG acquired by specifying 30 different 100 m/z segments. †: in these cases, the ratio was very large because the intensity of m+0 (numerator) remains consistent while that of m+2 (denominator) dropped due to the edge effect of the LIT. ‡: in these cases, the ratio was effectively zero or could not be defined because the intensity of either m+0 (numerator) or m+2 (denominator) was below detection limits.

3.3. Improved spectral binning method for intensity-independent noise filtering

Southam et al. [5] acquired the spectra as transient files and used MATLAB-based processing to obtain “full-profile” spectra with intact baseline and noise information. As indicated earlier, the noise from the “full-profile” spectra can be suppressed by conventional approaches [5, 8] but not the noise-truncated spectra such as those acquired from Thermo Orbitrap Fusion. The noise in these “reduced-profile” spectra can be removed using the recently reported intensity-independent method [9]. This method utilizes repetition rate filtering (RRF) which takes advantage of how often each peak appears within a small m/z interval (bin) across multiple scans. Although commonly used visual inspection-based noise thresholding can be applied to both types of spectra, it is impractical for S⁷ since each m/z segment has different noise levels. Considering these difficulties, we employed the RRF concept to determine the noise level of S⁷ spectra.

Spectral binning, necessitated by the RRF approach, requires the unique centroided m/z’s across scans to be put into a list in an ascending m/z order. We found that using the trend line of full width at half maximum (FWHM) as a criterion for the single-step bin width determination across the entire m/z range under ultra-high-resolution (R_{400 m/z} ≈ 370,000) acquisition produced imprecise bin boundaries (Figure 3). Also, in SIRM experiments, no assumptions can be made about the number of detectable isotoplogues or their intensity ratios, unlike in unlabled experiments where ¹³C_>5 ions would not be detectable for even very abundant metabolites of up to 50 carbons. Moreover, the detection of the ¹³C₀ isotopologue ions of metabolites is not assured (Figure S1c, [6]). Therefore the binning method needed to be as accurate as possible under ultra-high-resolution acquisition to avoid both false positives and negatives of stable istotope incorporation [4].

Figure 3. FWHM-based bin width threshold and averaging peaks under WFS acquisition.

Solvent mixture from 2.2.1 was injected and a single raw file was acquired via WFS with 5 microscans per scan spanning 10 minutes of acquisition time, up to 113 scans. (a) Skewed-overlay spectra from 8 selected scans out of the total 42 scans that detected signals in the m/z range of m+1 of PDMS [C₃₄H₁₀₂O₁₇ Si₁₇ + NH₄]⁺ across 113 scans. The scan numbers are annotated as #63, #27, #33, #28, #5, #23, #12, and #83 from top to bottom. The acquired raw files could be presented by either stick plots of profile data (black vertical dashed lines of 83^rd scan) or spectra that were generated by drawing the lines from point to point of the profile data. Centroided m/z values along with the profile data were available using the default centroiding algorithm (Thermo calls “label data”). These centroided m/z values were used in this study. The centroided m/z values are put next to the scan numbers and noted in the spectra as solid green lines. (b) The 113 scans were averaged via Xcalibur with the solid green line marking the centroided m/z value. Note that Xcalibur reports a single peak with the noise signals from the 83^rd scan merged into its base. (c) Distribution of the 43 peaks as the centroided m/z values from the 42 scans. Several scans had the exact same m/z values, which are indicated with colored dots (green and red); for instance, scans 28, 44, and 65 have m/z 1,277.35327. When the m/z range was binned based on FWHM, a putative single peak evident in (b) was separated into two (peaks 1 and 2) in (c) while merging the noise peaks into the analyte signals. The ~6 mDa bin width was calculated from the equation of a trend line of FWHM across m/z range of 150–1,600. The solid red lines from (a) to (c) show where the centroided m/z values of the noise peaks from the 83^rd scan are located in relation to the other peaks in the region.

In an effort to resolve the problem with FWHM-based binning in Figure 3, the difference between sequential m/z’s in the m/z list was obtained. These differences were found to be either equal to the minimum neighboring difference (MiND) or its integral multiples. The MiNDs were also observed to increase by a factor of 2 as the m/z range increased from 128–256 through 256–512 and 512–1,024 to 1,024–2,048 (Table S2), which our program took into account. A detailed example of how to determine the MiND is discussed in Table S2. The origin of this phenomenon is that digital computers store and do arithmetic on decimal numbers in binary floating-point format.

To determine the boundaries of bins, we inspected the maximum difference possible between adjacent m/z’s for given peaks of known contaminants (Figure 4), lipid standards (Figure S4), and prominent PC peaks from an unpolarized human macrophage (M0) extract identified based on the accurate mass (Figure S5, Table S3). Most of the known compound peaks were 1×, 2x, and/or 3× – 5×MiNDs apart from each other (Figures 4, S4). Also, most of the peaks identified as lipid isotopologue ions in the M0 extract showed narrow MiND distribution (1× – 5×MiNDs) as in Figure S5. The minimum distances we observed in this study were 6x or 7×MiNDs for two distinct, neighboring authentic peaks as in Figure 4j. This led to decision on creating a new bin each time the difference exceeds 5×MiND as the best practice for these samples acquired with ultra-high-resolution (R_{400 m/z} ≈ 370,000) and a moderate m/z range (150 – 1,600 m/z). Using this MiND-cognizant correction, we could accurately bin the PDMS peak (Figure 4f) whereas the FWHM-based approach and Xcalibur could not (Figure 3). For general practice, the bin boundaries should be user-defined and sample-, MS parameter-dependent.

Figure 4. BFP-based bin width threshold determination.

Distribution of the peaks with m/z values of: (a-f) m+1 of PDMS [C₃₄H₁₀₂O₁₇Si₁₇ + NH₄]⁺ and (g-k) m+2 of PDMS [C₅₀H₁₂₀O₂₁Si₁₇ + 2H]²⁺ across 113 scans in WFS range (150–1,600 m/z). (a, g) Differences in m/z values between the red dots are indicated above the distribution plot in terms of nxMiND (n is positive integer). Bins generated via (b, h) 1×MiND, (c, i) 2×MiND, (d, e, f, j) 3×, 4×, 5×, and 6×MiNDs, and (k) 7×MiND binning.

Once the size of each bin was determined, the number of peaks that fall into each bin across all scans was counted for RRF. In order to determine a repetition rate threshold that was less strict than that set by Schuhmann et al. [9], yet remained practical, we examined the repetition rate of m + n (n ≥ 2) isotopologue peaks of known contaminants (e.g. polyethylene glycol) in different m/z segments. The minimum repetition rate was in general above 5 % for WFS and 10 % throughout the segments of S⁷ spectra (data not shown).

However, there were some peaks with 4×, 5×, 6× and 7×MiNDs on the boundaries of the bins in both standard mixtures and M0 extracts (Figures 4, S4, S5). The influence of those peaks on the downstream RRF could vary depending on how strict the RRF threshold was set. For instance, at 5 % RRF threshold with either 3×, 4×, 5×, or 6×MiND binning, two peaks of m+2 PDMS ions (Figure 4j) were reported, which corresponds to the ions [C₅₀H₁₂₀ O₂₁²⁸ Si₁₆³⁰ Si + 2H]²⁺ and [C₄₉¹³CH₁₂₀ O₂₁²⁸Si₁₆²⁸Si + 2H]²⁺. The lighter and the heavier ions exhibited 97 and 15 % repetition rate, respectively. If the RRF threshold were as strict as [9] (70 %), only the lighter ion would be reported. The total number of detected signals in the bin (131) exceeded that of scans (113) with 7×MiND binning (Figure 4k), which is unrealistic, and the Python script would drop this bin from the final peak list. Most of the peaks separated by 6× or 7×MiND were found on the boundaries of the bins (Figure S4), and exclusion of these peaks did not affect the integrity of the bins at the given RRF threshold.

We used a 10 % repetition rate as the noise threshold for the S⁷ method, which removed more than ~ 85 % of signals detected in each segment (Figure S6). These settings can be used as default values but may need optimization for different sample types that have very different distribution of ions.

3.4. Application of S⁷ to SIRM lipid profiling of 3D human macrophage spheroids

3D cell cultures are increasingly preferred biological models that better reflect the microenvironment of cells in tissues than their traditional 2D counterparts [10, 30]. 3D cell cultures typically show lower proliferation rates than their 2D counterparts [10, 30], leading to less de novo lipid biosynthesis or ¹³C incorporation in SIRM experiments. This issue is exacerbated with human macrophage spheroids as they do not proliferate and the amount of materials that can be obtained from donors are very limited. As a result, many ¹³C isotopologue ions have relatively low intensities and fall below detection limits using WFS. By greatly enhancing the S/N, the S⁷ method can detect many more ¹³C isotopologue ions in SIRM studies, leading to improved biochemical interpretability of 3D cell cultures and human macrophage spheroids in particular.

The lipid extracts from M0 and M1 type human macrophage spheroids were prepared as described in 2.1.3 for comparing the performance of the WFS and S⁷ methods. The MiND-binning followed by 5 % and 10 % RRF was applied to WFS and S⁷, respectively, as determined in 3.3. Over the full range of 150–1600 m/z, we obtained ~0.38 ppm RMS error using the WFS method with AGC = 1 × 10⁵, which was improved to ~0.25 ppm RMS error using the S⁷ method (Table 2).

Table 2. Mass errors for S⁷ and WFS with AGC targets of 1 × 10⁵.

N_ic and N_me indicate the number of internal calibrants and m/z values used for mass error calculation, respectively. Av: average, RMSE: RMS error, MAE: max absolute error, and SDE: SD error.

Scan type	Macroph age subtype	Replicate	Nic	Nme	RMS error (ppm)	Max absolute error (ppm)	SDof error (ppm)	Av RMSE (ppm)	Av MAE (ppm)	Av SDE (ppm)
S7	M0	1	12	29	0.250	0.718	0.166	0.263	0.767	0.176
		2	12	29	0.243	0.791	0.164
		3	12	29	0.292	0.791	0.197
	M1	1	12	29	0.244	0.718	0.165	0.235	0.658	0.150
		2	12	29	0.230	0.608	0.137
		3	12	29	0.230	0.644	0.148
WFS	M0	1	2	31	0.433	1.088	0.453	0.391	1.011	0.411
		2	2	31	0.361	0.985	0.392
		3	2	31	0.374	0.955	0.390
	M1	1	2	31	0.343	0.822	0.363	0.369	0.972	0.404
		2	2	31	0.399	1.066	0.433
		3	2	31	0.364	1.011	0.414

S/N was overall improved (Figure S7) and the number of detected ¹³C isotopologue ions for the example [PC + H]⁺ species (Table S4, Figure 5a, b), increased considerably using the S⁷ method. We calculated the % ¹³C isotopic enrichment of two of the PC lipids, which are pulmonary surfactants, dipalmitoylphosphatidylcholine (DPPC) and palmitoyloleoylphosphatidylcholine (POPC) in Figure 5c, d. The improved detection of isotopologue ions was evident for the heavier isotopologue ions (e.g. >m+17 for DPPC in M0 data, >m+8–11 for POPC in M1 data) where they were undetected using WFS method but detected with good S/N using S⁷ method.

Figure 5. S⁷ increases S/N and detects many more ¹³C isotopologue peaks than WFS method in lipid data of ¹³C labeled human macrophage spheroids, especially at lower abundance.

The same lipid extract as in Figure S7 were analyzed. Peaks detected in only one of three spectra were removed from the final m/z list for the assigned glycerophospholipids. Numbers of detected and assigned (a) ¹³C₀ ions and (b) ¹³C_>0 ions of [PC + H]⁺. **P-value < 0.01 by paired Student’s t test. Error bars represent the standard error of the sample mean. ¹³C fractional enrichment distribution of (c) DPPC and (d) POPC after natural abundance correction [34]. Inset figures show relative intensities (before natural abundance correction [34]) of selected isotopologue ions of each PC species demonstrating a substantial S/N improvement. m+3, m+even number, and m+odd number represent respectively ¹³C incorporation into the glycerol backbone, fatty acyl chains, and glycerol+fatty acyl chains.

DPPC is a major pulmonary surfactant and has been shown to modulate inflammatory response of human monocytes [31–33], thus it is important to understand its metabolism in macrophages. Regardless of the acquisition methods used, the natural abundance (NA) corrected [34] % ¹³C distribution of DPPC and POPC showed a decreased % ¹³C enrichment in the fatty acyl chains in M1 versus M0 subtype with little or no changes in the extent of ¹³C incorporation into the glycerol backbone (Figure 5c, d). In fact, the entire PC class exhibited the same trend (data not shown), which is consistent with decreased fatty acid biosynthesis or PC degradation in LPS+IFNγ-activated M1 macrophages. This is in contrast to the notion that fatty acid biosynthesis is enhanced in M1 macrophages based on the overexpression of mitochondrial citrate carrier and ATP-citrate lyase, which together could promote cytoplasmic acetyl CoA production from citrate, which is required for fatty acid synthesis [35]. However, citrate-derived acetyl CoA is also required for histone acetylation [36]. Our direct measurement of lipid biosynthesis using ¹³C₆-glucose as tracer suggests that the synthesis of at least the PC class of lipids was not enhanced by M1 activation of human macrophage. Instead, reduced synthesis of DPPC can lead to DPPC depletion, which can promote inflammatory response in macrophages [32].

4. Conclusions

Recognizing the edge effect in both the LIT-[25, 26] and the QMF-based acquisition [27–29] is of great importance to properly perform SIM-spectral stitching. This issue is even more problematic for proper data analysis and interpretation of SIRM data [1–3, 15, 18], where preserving complex isotopologue patterns in stitched spectra is critical. We found that using QMF for ion isolation along with optimizing the SIM widths helped minimize the edge effect while greatly improving the S/N of stitched SIRM lipid data. Moreover, our MiND-cognizant binning approach solved the problem of inaccurate FWHM-based binning method. However, the threshold of RRF needs to be user-defined and experimental-design dependent. For example, for biomarker discovery purposes, one may want to remove low natural abundance isotopologue ions with low repetition rate. In contrast, in SIRM studies, one would need to optimize the detection of low-intensity, isotopically-enriched peaks that are important for biochemical interpretation. It is also impractical to use large numbers of standards to determine the lipid-species-specific repetition rate threshold, particularly for SIRM studies where the vast majority of isotopically-enriched standards are unavailable [37]. Thus, further efforts for determining the optimum RRF threshold by inspecting the repetition rate of m + n (n ≥ 2) isotopologue peaks of known contaminants is under way, with the goal of incorporating different threshold values for various sample settings.

When applying the S⁷ method to the challenging SIRM study of non-proliferating, low sample size human macrophages, we demonstrated the ability to detect very low abundant (< 0.5% ¹³C enriched), heavily labeled isotopologue ions (e.g. ¹³C_>17), which were undetected with the WFS method. Such higher sensitivity affords a much greater information throughput [6] without compromising the sample throughput, and has the potential to alter the biochemical interpretation. Lastly, the method is readily tailored to other types of samples that require wide range MS1 with high ion population density, different ion abundance silhouettes, and/or low S/N.

Highlights.

• Sample-specific segments were designed for the improved spectral stitching method

• Use of quadrupole mass filter-selected ion monitoring eliminates edge effect

• Binary floating point-based binning creates precise and resolution-independent bins

• Method was used on carbon-13 incorporated lipid profiling of human macrophage samples

• Method provides higher sensitivity and dynamic range compared with wide full-scan

Abbreviations:

AGC

automatic gain control

CE

cholesterol ester

COM

component object model

DG

diacylglycerol

DI nESI FTMS

Direct infusion nanoelectrospray Fourier transform mass spectrometry

DMEM

Dulbecco’s Modified Eagle Medium

DPPC

Dipalmitoyl phosphatidylcholine

FBS

fetal bovine serum

FWHM

full width at half maximum

LIT

linear ion trap

LPC

lysophosphatidylcholine

LPS

lipopolysaccharide

MiND

minimum neighboring difference

NS

NanoShuttles

PBS

Phosphate-buffered saline

PC

phosphatidylcholine

PE

phosphatidylethanolamine

POPC

Palmitoyloleoylphosphatidylcholine

PREMISE

Precalculated Exact Mass Isotopologue Search Engine

QMF

quadrupole mass filter

RRF

repetition rate filtering

S⁷

sample-specific silhouette segmented-scan spectral stitching

SIRM

stable isotope-resolved metabolomics

SM

sphingomyelin

TG

triacylglycerol

WB

whole blood

WFS

wide full-scan