A mass spectrometer that measures blood plasma sees a few proteins at enormous abundance and thousands of others near the noise floor. Each peptide ion leaves a characteristic fingerprint: a short comb of peaks, one per heavy-isotope variant, spaced by about $1.00335/z$ in mass-to-charge, where $z$ is the ion's charge. Speech has a closely matching structure. A voiced sound is a comb of harmonics whose spacing is the pitch, and the field of pitch tracking has spent decades learning how to find such combs when each individual peak is buried in noise. Pitch trackers succeed by pooling evidence across all teeth of the comb, across time, and across a narrowed search range set by a coarse first guess. They also explicitly check the gaps between teeth to avoid "octave errors," which in the mass spectrometry setting are the confusion between charge $z$ and $2z$.
We propose to transfer that recipe to a specific, unsolved case in data-independent acquisition (DIA) proteomics. DIA is a mode in which the instrument fragments every ion in a wide mass window at once. For many low-abundance peptides it yields convincing fragment evidence while the intact precursor appears "missing" from the survey (MS1) scan. DIAmeter explicitly identifies such peptides in plasma but treats the absent precursor as a loss. Our claim is that most of these precursors are not absent. They are below the per-peak noise floor, and a detector that does three things can recover them as a usable signal. It must pool counts over isotope teeth and retention-time scans under the correct Poisson noise model. It must anchor its search in time and mass using the fragment evidence. And it must subtract the combs of brighter overlapping peptides before searching again. The recovered precursor signal would then provide independent evidence that raises identifications at fixed false discovery rate and improves quantification of the lowest-abundance proteins.
Core Claim
The obvious version of this intersection is to fit an isotope comb to a spectrum by least squares, and that is already done. Slutzki et al. fit isotope templates by non-negative least squares or least absolute deviation with a hard threshold, and NITPICK does so with non-negative LASSO and a Bayesian information criterion stopping rule. Both are, in effect, the least-squares harmonic fitting of Bánhalmi et al. rewritten for mass spectra.
What neither does, and what we propose, is a detection test, not a regression. The test is designed for the regime where each isotope peak is a handful of ion counts. It has four parts:
- Poisson likelihood. It replaces the Gaussian noise assumption that NITPICK itself concedes breaks down at low counts.
- Pooling over retention time. It pools over retention time using an elution profile taken from co-eluting fragment ion chromatograms. This plays the role of the coarse first-pass estimate in the two-stage design of Drugman and Alwan.
- Inter-tooth penalty. It imports their inter-harmonic subtraction term, which penalizes energy halfway between teeth, to settle charge-doubling ambiguity.
- Closed loop. It wraps the test in an analysis-by-synthesis loop in the style of Pavlovets and Petrovsky, refining mass offset and peak width while subtracting brighter neighbors.
Concretely, for a candidate precursor at monoisotopic mass-to-charge $m$ and charge $z$, let $y_{t,k}$ be the observed counts at scan $t$ in the bin of isotope tooth $k$, located at $m + 1.00335k/z$. Let $b_{t,k}$ be the local background, $\pi_k$ the averagine isotope fractions, $e(t)$ the normalized elution profile from the fragment chromatograms, and $a \ge 0$ the unknown amplitude. The comb score is the Poisson generalized log-likelihood ratio
$$\Lambda(m,z) = \max_{a \ge 0} \sum_{t}\sum_{k} \Big[ y_{t,k}\log\Big(1 + \frac{a\,e(t)\,\pi_k}{b_{t,k}}\Big) - a\,e(t)\,\pi_k \Big],$$
from which we subtract the same statistic evaluated at the half-spacing offsets $m + 1.00335(k - \tfrac12)/z$. The subtraction does the job of the $E((k-\tfrac12)f)$ term in residual harmonic summation. A true charge-$2z$ ion places counts exactly there, so the penalty suppresses a wrong charge-$z$ call. The statistic grows with the total expected signal counts summed over teeth and scans, which is where the gain over per-scan, per-peak detection comes from. The amount of that gain is an empirical question we measure directly.
Prior Work and Its Roles
Three works motivate the problem.
- The missing-precursor phenomenon. DIAmeter documents it in plasma and scores matches by fragment p-values, coelution, and precursor intensity rank under Percolator re-ranking. It does not try to reconstruct the weak MS1 envelope. Our method adds exactly that missing feature.
- Fragment clustering. Dear-DIA XMBD clusters fragment chromatograms in a learned latent space and recovers many low-intensity peptides. However, its embeddings enforce no isotope-comb physics, and the authors note limitations with modified peptides.
- Dynamic range in practice. The Generation Scotland proteomic study shows that dynamic range limits biomarker discovery in real cohorts.
Two works constrain the design.
- Noise model. NITPICK's admission that a Gaussian noise model produces false peaks at low signal forces the Poisson formulation.
- Thresholding. Slutzki et al. argue that L1 penalties are unstable and that a hard threshold maps more cleanly to signal-to-noise. Our likelihood-ratio threshold is calibrated by target-decoy competition rather than by a hand-set penalty, which sidesteps their dispute.
Four works supply method pieces.
- Two-stage search. Drugman and Alwan provide coarse-to-fine bounding and the inter-harmonic penalty.
- Analysis-by-synthesis. Pavlovets and Petrovsky provide a closed loop over a unimodal cost, which we apply to mass offset and peak width.
- Least-squares refinement. Bánhalmi et al. provide evidence for such refinement: gross error falls from 4.4% to 3.2% on the Edinburgh database relative to the autocorrelation initializer.
- Feedback verification. The multi-talker tracker of Guan et al. shows that verifying tracks against the raw lag surface prevents drift between overlapping sources. This motivates re-checking each subtracted comb against the raw MS1 map before accepting the residual.
We deliberately do not adopt the HERB blind dereverberation loop of Nakatani et al.. Its need for thousands of utterances to learn an inverse filter is a poor fit for per-peptide chromatographic tailing, which the fragment-derived profile $e(t)$ already captures.
Research Program
The detector operates on candidate peptides proposed by a library-free DIA search.
First pass. Each candidate's fragment chromatograms give an apex time, an elution profile $e(t)$, and a predicted precursor mass and charge range. This narrows the MS1 search to a window of a few scans and a few tens of ppm. This is the speaker-specific band of the two-stage pitch trackers.
Second pass. We compute $\Lambda$ with the half-spacing penalty for each allowed charge. We then refine the mass offset and an isotope-peak width by golden-section search on the likelihood, keeping the best charge.
Interference loop. Before scoring a trace candidate, we fit and subtract the combs of all higher-scoring overlapping precursors in the same window, using Poisson-consistent residuals. After subtraction, we verify against the raw map that no subtracted comb has drifted onto the target's teeth, and then iterate once or twice.
Output. The final statistic, its fitted amplitude, and the correlation between the precursor time trace and $e(t)$ enter Percolator as new features alongside the existing ones.
The background $b_{t,k}$ is estimated from nearby off-comb bins. For TripleTOF data, intensities approximate ion counts. For Orbitrap data, we convert intensities to approximate charge counts with a per-run noise calibration. We report the two instrument families separately because the Poisson assumption is cleaner on the first.
Baselines
Naive reference. The do-nothing reference is the same library-free search with no new MS1 feature: DIAmeter's scoring as published, including its precursor-intensity-rank feature, re-ranked by Percolator. Any gain must show up as more peptides at the same 1% peptide-level false discovery rate against this exact pipeline. We will run on the same files, the same database, and the same Percolator settings.
Strongest physics-based competitor. This is non-negative template regression applied to the same retention-time-anchored MS1 windows. We run Slutzki et al.'s NNLS with a hard coefficient threshold, and NITPICK's non-negative LASSO with BIC stopping. Both are run per scan and on retention-time-summed spectra, so that the only differences from our method are the noise model, the inter-tooth penalty, and the subtraction loop. Their outputs enter Percolator as features in the same way ours do. Each method gets the same tuning budget on a held-out run. These are the closest existing methods, because they already exploit the comb. Slutzki et al. report significant precision and recall gains at 10 fmol on MALDI-TOF, but the file gives no numeric values.
Strongest end-to-end system. This is Dear-DIA XMBD. On HYE124 it reports identifying 12.3 times more unique peptides than DIA-Umpire, and it reports more peptides and proteins than Spectronaut 14. We will run its released model on HYE124 and LFQbench. The comparison of interest is not a replacement but a union: whether our MS1 feature, added to its candidate list, raises identifications at fixed false discovery rate. Beating it means a statistically reliable increase in peptides at 1% FDR on held-out runs with no loss in entrapment-estimated false discovery.
Ablations. The ablations isolate each ingredient by removing it in turn:
- Gaussian versus Poisson likelihood.
- With versus without the half-spacing penalty, scored by charge-assignment accuracy.
- With versus without fragment-anchored $e(t)$, replaced by a free Gaussian elution shape.
- With versus without the subtraction loop.
The key claim lives in the combination of Poisson pooling and fragment anchoring. If the Gaussian NNLS on retention-time-summed spectra matches the full method, that claim fails.
Why not just…?
Why not just run NNLS or NN-LASSO template matching on retention-time-summed MS1? This is the real threat. Summing over scans already pools evidence, and averagine templates already encode the comb. Our argument is that at a few counts per tooth, least squares weights bright noise spikes from co-eluting ions as heavily as signal and produces false peaks, as NITPICK concedes. Summing also discards the elution shape that separates a true precursor from a neighbor eluting a few scans away. The Gaussian-versus-Poisson and anchored-versus-free ablations on the thinning benchmark below decide this directly. If the gap is within noise, we will report that the contribution reduces to "feed retention-time-summed NNLS into Percolator," which is still useful but much smaller.
Why not just trust fragment evidence and ignore MS1, as DIAmeter already does successfully? Fragment evidence alone is strong, but fragments of co-isolated peptides in a wide window are shared and confounded. An independent precursor measurement adds orthogonal evidence that Percolator can exploit, and it also gives a quantity for quantification that is less contaminated than shared fragments. If the Percolator feature weights for our score come out near zero and identifications do not rise, the alternative wins.
Why not just let a learned model such as Dear-DIA XMBD absorb the MS1 signal? A learned model could in principle discover comb pooling. But Dear-DIA's published design clusters fragment chromatograms and does not model precursor isotope physics, and it reports difficulty with modified peptides. An explicit Poisson comb test needs no training data and applies unchanged to any composition the averagine model covers. The union experiment and a breakdown of gains by peptide abundance and modification status show whether the explicit test adds anything beyond the learned system.
Why not just preconcentrate or amplify the sample physically? Seed amplification such as RT-QuIC works only for self-propagating aggregates. Ion concentration polarization and selective nano-adsorbents enrich chosen targets rather than the untargeted proteome. These approaches are complementary to computation, not substitutes for it: any enriched sample still ends in a spectrum with a noise floor.
Proposed Experiments
Ground-truth detection benchmark. The decisive check uses binomial thinning, which keeps Poisson counts exactly Poisson. From TripleTOF HYE124 runs, we take precursors confidently identified in both MS1 and MS2. We then thin only their precursor counts and fragment counts by factors from 1/10 to 1/1000 and leave the real plasma-like or lysate background untouched. This produces trace precursors at known positions inside real interference. We plot detection probability against injected expected counts at a fixed false-alarm rate, estimated from decoy comb positions with shifted spacing. We compare the full method, every ablation, and the regression baselines. Success means the full method reaches a given detection probability at substantially fewer injected counts than retention-time-summed NNLS, and that charge-assignment errors fall when the half-spacing penalty is on.
Benchmark identification and quantification. On HYE124 and LFQbench, we count peptides and proteins at 1% FDR and validate the false discovery rate with an entrapment database of a species absent from the mixture. We also measure log-ratio error of yeast and E. coli proteins against their known mixing ratios, stratified by abundance decile. The prediction is that gains concentrate in the bottom deciles. On the TripleTOF plasma data used by DIAmeter, we measure how many peptides previously lacking MS1 evidence acquire a significant precursor score. No reported result was found on Lacuna for DIAmeter's per-dataset counts there, so this baseline will be re-run, not quoted.
Cross-instrument check. We repeat the identification analysis on OxMetYeast across its 20, 10, and 5 m/z isolation widths. This tests whether the benefit grows with window width, as interference does. It also tests whether the Orbitrap count calibration preserves the Poisson advantage. If it does not, the method should be scoped to ion-counting instruments.
References
- DIAmeter: matching peptides to data-independent acquisition mass spectrometry data
- Isotope pattern deconvolution for peptide mass spectrometry by non-negative least squares/least absolute deviation template matching
- NITPICK: peak identification for mass spectrometry data
- Fundamental frequency estimation by least-squares harmonic model fitting
- Joint robust voicing detection and pitch estimation based on residual harmonics
- Robust HNR-based closed-loop pitch and harmonic parameters estimation
- Dear-DIA XMBD : Deep Autoencoder Enables Deconvolution of Data-Independent Acquisition Proteomics
- Proteomic biomarkers of cognitive function, APOE ε4 status, and dementia in Generation Scotland
- Multi-Pitch Detection for Co-Channel Speech Utilizing Frequency Channel Piecewise Integration and Morphological Feedback Verification Tracking
- One Microphone Blind Dereverberation Based on Quasi-periodicity of Speech Signals
- A single ultrasensitive assay for detection and discrimination of tau aggregates of Alzheimer and Pick diseases
- Folding-paper-based preconcentrator for low dispersion of preconcentration plug
- A porous nano-adsorbent with dual functional groups for selective binding proteins with a low detection limit