A data-dependent run decides, in real time and one scan at a time, which precursors are worth fragmenting. That decision is where most of the sample goes missing. In standard LC runs of a cell lysate analysed by high-resolution MS, more than 100,000 likely peptide features were detected, of which only about 16% had been targeted for MS/MS, and the median precursor ion fraction — the share of the isolation window's ion current belonging to the intended precursor — was 0.14 [1]. Data-independent acquisition removes the decision. Instead of selecting precursors, the quadrupole steps through a fixed ladder of isolation windows and fragments everything in each one, on every cycle, for the whole gradient. Data-dependent acquisition and the workflow built around it are treated separately in the LC-MS/MS proteomics guide; this article is about what changes when precursor selection is taken out.
The original SWATH implementation cycled repeatedly through 32 consecutive 25-Da precursor isolation windows spanning 400–1200 m/z on a fast, high-resolution quadrupole–quadrupole time-of-flight instrument, generating time-resolved fragment ion spectra for all analytes detectable in that precursor range [2]. Because acquisition is unconditional, a peptide can be interrogated from its fragment traces even when its precursor was never visible in a survey scan, and that report showed query peptides confidently identified across a dynamic range of four orders of magnitude, quantified with a consistency and accuracy comparable with selected reaction monitoring [2]. That is the trade in one sentence: you give up precursor selectivity at acquisition and buy it back computationally, or you do not get it back at all.
What data completeness buys, and what it does not
A practical reason to move to DIA is the missing-value rate. On a controlled sample set, a DIA method implemented on an Orbitrap platform with retention-time-normalised (iRT) spectral libraries and targeted data extraction gave reproducibility of peptide detection above 98%, yielding quasi-complete data sets, against 49% for shotgun proteomics on the same comparison [3].
Read that number for what it is. It states how often the same peptide is reported across repeated measurements, not how often it is reported correctly. A matrix with few holes is far easier to feed into ratio tests, clustering and regression than one riddled with them, and it removes an entire class of imputation decisions from the analysis. It does not certify that the values filling those cells are free of contribution from co-eluting ions. Completeness and quantitative correctness are separate axes, and a low missing-value rate is not evidence about the second.
Chimeric spectra are the core analytical problem
Every DIA fragment spectrum is a sum. Widening the isolation window from the narrow selection of a targeted method to tens of daltons means product ions from many co-fragmenting precursors land in one spectrum with nothing marking which came from where. The cost has been quantified: DIA data are noisy owing to a typical five- to tenfold reduction in precursor selectivity compared with data obtained by data-dependent acquisition or selected reaction monitoring [4]. Two families of remedy exist at the acquisition stage. Multiplexing recombines several narrow windows per spectrum and demultiplexes afterwards; the MSX scheme increased precursor selectivity fivefold [4]. Alternatively, spectra collected with overlapping rather than adjacent or random windows and then computationally demultiplexed improved precursor selectivity by nearly a factor of two, with no loss in mass range, mass resolution, chromatographic resolution or scan speed, and in a six-protein bovine mix spiked into a yeast background this yielded a 64% improvement in sensitivity and 17% more peptides detected [5].
Both results say the same thing about your data. The window width you chose sets a ceiling on how well any downstream algorithm can attribute a fragment to a precursor. No scoring function recovers information the quadrupole never separated.
Window schemes: variable widths and the cycle-time trade
Uniform windows waste duty cycle, because precursor density is not flat across the mass range. Variable-width schemes size each window to equalise the load. On a trapped-ion-mobility platform, windows placed optimally according to precursor density in the m/z and ion-mobility plane, combined with in-depth project-specific libraries and one specific liquid chromatography system, reproducibly identified over 7,700 proteins from a human cell line in 44 minutes with quadruplicate single-shot injections, consistently quantified more than 6,000 proteins in mammalian cell lysates on 11-minute gradients, and quantified more than 35,000 phosphosites in triplicate 21-minute runs on a stimulated human cell line [6].
The constraint bounding every window scheme is cycle time. Each window costs one MS/MS scan, and the full ladder must complete often enough to sample each chromatographic peak repeatedly, or quantitative precision degrades however selective the individual spectra are. Narrower windows mean more of them, which means either a longer cycle or a faster analyser. The narrow-window approach resolves this by scan rate: high-resolution MS1 scans paired with MS/MS at roughly 200 Hz using 2-Th isolation windows on an asymmetric track lossless analyser reached a depth of about 10,000 human protein groups in half an hour and about 7,000 proteins in five minutes — reported as threefold higher coverage than the state of the art at the time — and quantified more than 14,000 protein groups in a single half-hour run of a three-species proteome mixture [7]. Without that scan speed, the same 2-Th ladder would simply undersample the peaks.
Setting a scheme is therefore a three-way reconciliation, and the order of the steps matters. Fix the gradient and the expected chromatographic peak width first; divide by the number of points you require across a peak to obtain the longest tolerable cycle; divide that by the accumulation or transient time your analyser needs per MS/MS scan to obtain the number of windows you can afford. Only then decide how to distribute those windows across the precursor range, uniformly or by density [6]. If the resulting widths are too coarse for the complexity of the digest, the remaining moves are acquisition-side selectivity gains — the multiplexed MSX scheme [4], or overlapping windows, which improve selectivity with no loss of scan speed [5] — or a faster analyser [7]. Running the sequence backwards, choosing a window width first and letting the cycle stretch to accommodate it, is what produces DIA data that identify well and quantify badly.
Spectral library versus library-free analysis
DIA analysis is peptide-centric. Rather than asking what a spectrum is, you ask whether a nominated peptide is present and score the evidence for it, which requires knowing in advance what to look for. The classical route supplies that knowledge as an experimental spectral library of fragment ion spectra, mined against the complete fragment ion maps [2].
Library-free analysis removes the dependency in one of two ways. Pseudo-spectrum approaches deconvolute the acquisition first: DIA-Umpire detects precursor and fragment chromatographic features and assembles them into pseudo-tandem MS spectra that conventional database-searching and protein-inference tools can identify, allowing sensitive untargeted analysis without a spectral library, then performs targeted extraction of quantitative information for peptides initially identified in only a subset of the samples, which makes quantification more consistent across the set [8]. Direct peptide-centric detection skips the pseudo-spectrum altogether: because DIA tandem spectra are often highly multiplexed with product ions from multiple co-fragmenting precursors, PECAN reports evidence of detection from product ion scoring alone, which enables detection of low-abundance analytes with poor precursor ion signal [9]. Chromatogram libraries occupy the middle ground, capturing fragment ion chromatographic peak shape and retention time for every detectable peptide and using those coordinates to calibrate a protein database or an existing spectrum library to one specific mass spectrometer and chromatography setup; on human and yeast cells this detected 20–25% more peptides than data-dependent-acquisition-based spectrum libraries alone [10]. On top of either route, learned scoring and signal correction are now standard: DIA-NN exploits deep neural networks with new quantification and signal-correction strategies, with interference correction as its stated mechanism, and is particularly beneficial for high-throughput work on fast chromatographic methods [11].
The choice is not neutral for error control. An experimental library restricts the search space to peptides for which you already hold evidence in comparable material. A predicted, proteome-wide in silico library does not.
Why FDR in DIA is not the DDA problem
In data-dependent work, target–decoy competition is organised per spectrum: each spectrum yields one best match and decoys compete for it. Peptide-centric DIA has no such structure. Each query peptide is scored against the whole run, so the number of tests is set by the library rather than by the number of spectra, and a genuinely absent peptide still returns a best-scoring peak group somewhere. Statistical concepts developed for spectrum-centric discovery proteomics can be adapted to large-scale DIA analysed by peptide-centric scoring, but optimal trade-offs between sensitivity and specificity require careful consideration of the relationship between the proteins in the sample and the proteins represented in the spectral library, and a global analyte constraint is proposed to prevent the accumulation of false positives across large-scale data sets [12]. The same work argues that well-established confidence criteria should be reported separately for detected peptide queries, for peptides and for inferred proteins, because these are different quantities [12].
How well do current tools meet that standard? Entrapment procedures, which expand the search space with verifiably false entries, are the standard validation. A systematic theoretical treatment of them identified three prevalent methods for validating FDR control in use — one invalid, one providing only a lower bound, and one valid but under-powered — and, applying a more powerful evaluation, found that no DIA search tool consistently controls the FDR, with particularly poor performance on single-cell datasets [13]. A separate ground-truth evaluation, using recombinant proteins synthesised from full-length human cDNA libraries and searched library-free, measured average FDRs across the synthesised protein mixtures of 0.538%, 0.389% and 0.385% at precursor level for three successive DIA-NN versions, against 2.85%, 1.81% and 1.81% at protein level in the same comparison [14]. Note the shape of that result rather than the absolute values: protein-level error ran several-fold higher than precursor-level error in every version, and the newer versions improved control by adopting a more conservative identification approach [14]. Controlling FDR is challenging with in silico libraries precisely because of the vast amount of peptide information they contain [14].
An entrapment result tells you whether a pipeline's reported error rate was honest on the data it was run against. It does not transfer. A tool that controls the FDR on a bulk lysate benchmark is not thereby validated for a low-input experiment, and single-cell data, the extreme of low input, is where that systematic evaluation found control weakest [13]. The same caution applies to library provenance: because the sensitivity–specificity trade-off depends on how the sample's protein content relates to the library's contents, an FDR validated with one library says little about the same tool run with a proteome-wide predicted library [12, 14].
Two operational consequences. Do not report a protein-level DIA result at the error rate you set for precursors. And do not assume the figure your software prints has been validated on data resembling yours, particularly at low input.
Retention-time alignment and interference-aware scoring
Peptide-centric scoring has to know where in the gradient to look. Indexed retention time supplies the anchor: an empirically derived, dimensionless, peptide-specific value fixed relative to a standard set of reference peptides and transferable across laboratories and chromatographic systems, which permitted acquisition windows more than four times smaller than in silico retention time predictions when it was introduced for scheduling targeted measurements [15]. DIA libraries adopted the same normalisation, and the quasi-complete matrices described above were produced with iRT-normalised libraries [3].
Anchoring is not alignment. Chromatography drifts across a batch, and a peak group correctly picked in the first injection may be picked from a neighbouring feature in the fortieth. Cross-run alignment that uses fragment ion data to enforce consistent peak-picking and quantification reduced the identification error by more than threefold at constant recall, relative to a state-of-the-art SWATH-MS analysis without alignment, while correcting for highly nonlinear chromatographic effects [16]. Alignment is therefore part of the error model, not cosmetic post-processing.
Interference-aware scoring is the other half of the same problem. Every extracted fragment trace may carry signal from a co-isolated precursor, so a quantifier that sums transition areas without checking inherits that contamination; wide isolation windows limit the achievable accuracy of quantification as well as identification [5]. Demultiplexing at acquisition [5] and signal correction at analysis [11] attack the same target from opposite ends. When inspecting a peak group, treat identification confidence and quantitative accuracy as separate verdicts: co-elution of fragment traces with matching peak shape and library-consistent relative intensities is the evidence for the second, and a single contaminated transition can distort a ratio in a peak group you are right to call detected.
What a DIA quantitation licenses you to conclude
A DIA run gives you a re-mineable fragment ion map. The acquisition is unbiased with respect to the questions you have not asked yet, and the same file supports iterative re-mining and dynamic extension of protein probing after the fact [2]. It gives you far fewer missing values than shotgun acquisition on comparable material [3]. It does not give you a per-spectrum assignment, and it does not give you calibrated protein-level error rates for free [12, 13, 14].
Three things belong in the record if the result will be reused. First, the window scheme — count, widths, whether overlapped or multiplexed, and cycle time — because both selectivity and points across the chromatographic peak follow from it [4, 5, 6, 7]. Second, the library: experimental, chromatogram-calibrated or fully in silico, and the search space it defines, since that governs what the reported FDR means [10, 12, 14]. Third, the alignment and scoring configuration, including whether a global analyte constraint was applied across runs and what was done about interference [11, 12, 16]. Two analyses of the same raw file under different libraries are not the same measurement, and a peptide list from one is not directly comparable with a peptide list from the other unless both are stated.