MSPLIT-DIA: Sensitive Peptide Identification for Data Independent Acquisition

To the Editor

Recently developed data-independent acquisition (DIA) approaches for mass spectrometry data collection are gaining traction in the proteomics field. We present MSPLIT-DIA (Mixture-Spectrum Partitioning using Libraries of Identified Tandem mass spectra) as a spectral matching tool for untargeted and sensitive peptide identification in DIA data (http://proteomics.ucsd.edu).

Despite its sensitivity on modern mass spectrometers, the semi-stochastic nature of Data-Dependent Acquisition (DDA) leads to sampling a different subset of peptides each time a sample is analyzed, resulting in missing peptide identifications and decreased reproducibility across multiple runs. DIA strategies aim to alleviate this problem by systematically isolating and fragmenting ions based only on their m/z and not their intensities. DIA strategies often segment the usable m/z range into wide isolation windows (e.g. 25 Da windows in SWATH¹), generating complex spectra with multiple peptides that cannot be readily identified with DDA tools. Instead, DIA data analysis tools are mostly based on targeted extraction of quantitative information using SRM-inspired strategies¹, though recent approaches published in Nature Methods, one of which³ was under review concurrently with this submission, extract pseudo-MS/MS spectra which are then searched with DDA database search tools^{2, 3} Nevertheless, computational tools that explore alternative strategies for identifying peptides in multiplex spectra are still needed.

We introduce MSPLIT-DIA (Mixture-Spectrum Partitioning using Libraries of Identified Tandem mass spectra), a spectral matching tool for untargeted peptide identification in DIA data (Fig. 1a; Supplementary Methods). Due to the likely presence of many peaks from co-eluting peptides in each multiplexed spectrum, MSPLIT-DIA uses spectrum projections to match library spectra to each DIA spectrum, and spectrum-spectrum match (SSM) similarity is then evaluated using the normalized dot product. Since peptides are acquired and analyzed throughout their elution profile, MSPLIT-DIA also evaluates the similarity of the matched peaks between library spectra and multiplexed spectra across multiple consecutive DIA spectra. Finally, the statistical significance of SSMs matches is assessed at 1% peptide-level false discovery rate (FDR) using the target-decoy approach. For each multiplexed spectrum, all SSMs with FDR≤1% are returned as matches. MSPLIT-DIA effectively identified up to 10 peptides per spectrum, with complex samples such as human lysate generating a predominance of spectra containing more than one peptide (Fig. 1b; Supplementary Fig. 1).

Fig. 1. MSPLIT-DIA identification of peptides, proteins and protein-protein interactions.

(a) Overview of the MSPLIT-DIA identification process (see also the Supplementary Note for details). (b) Peptide multiplicity as identified by MSPLIT-DIA in multiplexed spectra from DIA runs of varying complexity. (c) A human lysate was analyzed by itself (left) and with a spiked-in standard 48-protein mixture (UPS1, right); the number of unique peptides identified using six different data analysis approaches is shown. MSPLIT-DIA analysis was performed either with a library from paired Data Dependent Acquisition (DDA) runs of the same sample (green) or using the large generic SWATH-Atlas library (purple). (d) Number of peptides detected across four runs (top panel) and reproducibility analysis comparing MSPLIT-DIA and MSGFDB-DDA (bottom panel) in relation to peptide abundance (x-axis). (e) Comparison of MSGFDB-DDA and MSPLIT-DIA methods as applied in a semi-quantitative approach to detect protein-protein interactions from affinity-purified samples from two bait proteins (EIF4A2, MEPCE) and a negative control (GFP). (f) DIA analysis workflow using MSPLIT-DIA circumvents the need to use additional DDA runs, spike-in peptides or manual curation for generation of sample-specific assay libraries. (g) Results of PeakView and MSPLIT-DIA peptide quantification with or without retention-time alignment and with or without an MSPLIT-generated assay library.

Since DDA is the current standard for sensitive peptide identification, we compared MSPLIT-DIA with MSGFBD⁴ analysis of DDA data (Fig. 1c, Supplementary Figs. 2–4). While the performance was comparable when using spectral libraries generated from the same samples, MSPLIT-DIA with the SWATH-Atlas⁵ spectral library identified 26–31% more human peptides than the corresponding DDA analysis. MSPLIT-DIA also identified 66–89% more human peptides than DIA-Umpire² and 81–88% and 86–107% more than PeakView⁶ and Skyline, respectively, in the same DIA runs. MSPLIT-DIA further enabled much more reproducible observations across 4 runs than DDA (Supplementary Fig. 5; 70% versus 50% at 1% FDR). The reproducibility gains were most pronounced for the 60% lower-abundance human peptides (Fig. 1d, Supplementary Fig. 5; 59% gains). This is important for comparative studies, where biological conclusions are drawn from the detection and non-detection of low-to-medium abundance peptides and proteins across samples.

We benchmarked MSPLIT-DIA for the analysis of protein-protein interactions (a major application of comparative proteomics), by re-analyzing DDA and DIA data for biological triplicates of the baits EIF4A2 and MEPCE compared to the negative control GFP⁶ (Supplementary Figs. 6,7). Since MSPLIT-DIA used in an untargeted manner yields spectral counts (Fig. 1d) that are not biased by precursor ion selection as in DDA dynamic exclusion protocols, we used these as a rough measure of abundance. The substantial ~3–4× gains in spectral counts synergized with the reproducibility improvements to yield better signal-to-noise for approaches relying on spectral counts, such as SAINT⁷. For MEPCE and EIF4A2, this resulted in the confident detection of ~33% more interacting proteins (Fig. 1e) that are consistent with the biological function of the bait proteins (Supplementary Tables 1,2; Supplementary Fig. 7). Thus, even when only coupled to rough abundance measurements, the sensitivity, reproducibility, and spectral count increases obtained with MSPLIT-DIA improve the sensitivity of detection of interactions. Since the generic SWATH-Atlas library⁵ was even better at detecting interactors than the sample-specific library (Supplementary Figs. 6,7; and more sensitive than DIA-Umpire² for protein identifications on these samples), our results also suggest that time-consuming generation of tailored libraries or addition of external retention time (RT) calibration standards may become superfluous as more spectral libraries become publicly available.

While we contrasted above the sensitivity of targeted extraction tools with MSPLIT-DIA, these are in fact complementary approaches (Fig. 1f, Supplementary Fig. 8). Although targeted extraction tools performed relatively well on libraries generated from the paired DDA samples (with matched complexity, instrument parameters and chromatographic resolution; Supplementary Fig. 2), this was not the case with the large generic SWATH-Atlas library⁵ (Fig. 1g; Supplementary Fig. 9). First, we showed that MSPLIT-DIA greatly facilitated targeted extraction by assisting in RT alignment without the need for spike-in standards as peptides identified by MSPLIT-DIA in the DIA run served as markers for alignment (“MSPLIT-DIA-assisted RT alignment” in Fig. 1g and Supplementary Fig. 9). Second, restricting the targeted quantification search space to only MSPLIT-DIA-identified peptides yielded much smaller assay libraries that either enabled (Skyline) or systematically improved (PeakView⁶, OpenSWATH¹) targeted extraction results (“MSPLIT-DIA library” in Supplementary Fig. 9). Altogether, these processes substantially simplified the targeted extraction of quantitative data for up to 88% of the peptides identified by MSPLIT-DIA without affecting the reproducibility in the quantification of these newly identified peptides (Supplementary Fig. 10). Assay libraries for targeted extraction tools are automatically generated by MSPLIT-DIA to facilitate coupling of sensitive identification with accurate quantification from DIA data.

Abbreviations

MS/MS

Tandem mass spectrometry

DIA

Data independent acquisition

DDA

Data dependent acquisition

MSPLIT

Mixture-Spectrum Partitioning using Libraries of Identified Tandem mass spectra

PSM

Peptide-Spectrum Match

SSM

Spectrum-Spectrum Match

FDR

False Discovery Rate

SRM

Selected Reaction Monitoring