Calibrated Unfolding with Conserved-Sequence Priors for Peptide Fragments Below the Noise Floor

Every mass spectrometer has a noise floor, and every low-abundance protein in a clinical sample — a viral structural protein, a circulating damage biomarker — produces peptide fragment ions whose intensities sit uncomfortably close to it. When a faint fragment peak is discarded as noise, the protein goes unquantified, no matter how good the downstream search software is. Particle physicists faced an almost identical problem in neutron spectroscopy and low-energy recoil detection, and solved it not by building quieter detectors but by mathematically inverting a calibrated detector response: they characterize exactly how a true signal of known energy gets smeared and attenuated by the instrument, then iteratively unfold measured pulse-height distributions back into true spectra, recovering information down to thresholds the raw data could not reach (High-fidelity unfolding of 252Cf neutron spectra in EJ-301 scintillators via calibrated response matrices and Monte Carlo simulations). Proteomics has never done this. Meanwhile, a separate literature on conserved region targeting shows that essential pathogen and host proteins contain short sequence motifs that evolution cannot change without breaking function — invariant anchors like the eight-amino-acid VP1 epitope WFYDGYPT shared across Enterovirus A serotypes (Discovery of a broad-spectrum monoclonal antibody recognizing a conserved, linear epitope WFYDGYPT on VP1 protein of Enterovirus A species) or the RNA-dependent RNA polymerase cores conserved across divergent astroviruses (Non-Human Primates Harbor Diverse Mammalian and Avian Astroviruses Including Those Associated with Human Infections).

This proposal combines the two. The core claim: if we build an instrument-calibrated response matrix for tandem mass spectra — the mass spectrometry analogue of the Gaussian energy broadening and Monte Carlo response calibration used in neutron spectroscopy — and use conserved sequence motifs as a physically motivated prior inside the iterative unfolding, we can recover and accurately quantify peptide fragment ions whose measured peaks are below the conventional signal-to-noise threshold, extending the effective dynamic range of existing instruments for exactly the conserved, low-abundance targets that matter most for pathogen surveillance and clinical prognosis.

Core Claim

Three facts from the supplied literature make this both plausible and untested. First, calibrated response-matrix unfolding demonstrably recovers spectral information near the detection baseline in a physically analogous setting: the EJ-301 scintillator work reconstructed true neutron spectra down to a 0.35 MeV threshold with average deviation below 4.28% and best agreement of 1.42%, using Gaussian energy broadening parameters calibrated from gamma-source Compton edges and Maximum Likelihood Expectation Maximization (MLEM) inversion (High-fidelity unfolding of 252Cf neutron spectra in EJ-301 scintillators). Second, mass spectrometry's own machine-learning tools treat spectra as raw token sequences or binned intensities and never model the instrument's transfer function: the de novo sequencer AdaNovo fights noise with a dual-decoder information-theoretic architecture rather than a detector model (AdaNovo: Towards Robust De Novo Peptide Sequencing in Proteomics against Data Biases), the DIA deconvolution tool Dear-DIA XMBD uses a black-box autoencoder with no physical grounding (Dear-DIA XMBD: Deep Autoencoder Enables Deconvolution of Data-Independent Acquisition Proteomics), and the fragment-probability model Pep2Prob learns fragmentation statistics purely empirically (Pep2Prob Benchmark: Predicting Fragment Ion Probability for MS2-based Proteomics). Third, in the low-energy collision regime the intensities of weak fragment ions are governed by the same kind of threshold-limited momentum-transfer physics that dark matter detectors confront — where simple models can fail theoretically yet match rigorous calculations numerically, as the phonon-mediated Migdal literature shows the free-recoil approximation matching full quantum field theory down to 10 MeV despite breaking down in principle below 50 MeV (Phonon-mediated Migdal effect in semiconductor detectors). That tension is a caution and an opportunity: a simple calibrated response model for the mass spectrometer may be numerically adequate far below where naive theory says it should be, but this must be validated, not assumed.

The specific claim we will test: a response matrix $\mathbf{R}$ mapping true fragment intensities to measured spectra, calibrated on synthetic peptide standards and inverted with a conservation-informed MLEM prior, recovers fragment ions at signal-to-noise ratios where standard peak picking, empirical transformers, and autoencoder deconvolution all fail, while producing calibrated absolute intensities that none of those methods provide.

Research Program

The program has three components. The first is construction of the response matrix. For a given instrument and collision mode, entry $R_{ij}$ is the probability that a fragment ion whose true mass-to-charge ratio and intensity fall in bin $j$ is recorded in measured bin $i$. This combines a Gaussian mass-accuracy broadening term (the direct analogue of the Gaussian Energy Broadening used in scintillator calibration), a low-intensity detection-efficiency term that falls toward the electronic noise floor, a saturation term at high intensity, and a low-mass cutoff. Each term is calibrated empirically, exactly as the neutron work calibrated broadening from $^{137}$Cs and $^{22}$Na Compton edges: here, synthetic peptide libraries with known fragment identities and controlled dilution series provide the calibration standards, and the PROSPECT dataset — 838,000 unique peptides, 61.7 million labeled spectra, and 5.7 billion annotated fragment peaks across 26 characterized neutral-loss types (PROSPECT: Labeled Tandem Mass Spectrometry Dataset for Machine Learning in Proteomics) — is the natural source, because synthetic peptides isolate fragmentation physics from biological messiness.

The second component is conservation-informed unfolding. Given a measured spectrum $\mathbf{y}$, we invert

$$\mathbf{y} = \mathbf{R}\,\mathbf{x} + \boldsymbol{\epsilon}$$

where $\mathbf{x}$ is the vector of true fragment intensities and $\boldsymbol{\epsilon}$ is noise, using the MLEM update iterated from an initial estimate $\mathbf{x}^{(0)}$ that is not uniform but weighted toward the fragment series (b- and y-ions, with their neutral losses) expected from peptides containing a known conserved motif. This is where conserved region targeting earns its place: for a surveillance or biomarker assay, we already know which invariant anchors we are hunting — the WFYDGYPT epitope of Enterovirus A, the invariant pK205R protein of African swine fever virus that current assays detect only serologically (A highly efficient indirect ELISA and monoclonal antibody established against African swine fever virus pK205R), or the matrix-gene-coded proteins that anchor pan-subtype influenza detection (Identifying Influenza Viruses with Resequencing Microarrays). The prior does not assume the peptide sequence; it restricts the unfolding's degrees of freedom to fragment patterns consistent with any peptide containing the motif, which is what stabilizes the inversion when most true peaks are individually sub-threshold. This is the same regularization logic that makes iterative unfolding stable near baseline in neutron spectroscopy, where unregularized inversion produces severe numerical artifacts.

The third component is honest scope-setting. Unfolding cannot fix what the instrument never acquired: it recovers fragment-level information within recorded spectra, so the primary application regime is data-independent acquisition windows and targeted scans where low-abundance peptides are fragmented but their fragments drown near the noise floor — not stochastic data-dependent acquisition, which never selects those precursors at all. We also note that the existing closest work combining conserved targets with mass spectrometry — cross-linking MS on the conserved domains of S. pyogenes M1 protein (Structural determination of Streptococcus pyogenes M1 protein interactions with human immunoglobulin G using integrative structural biology) — uses standard high-energy fragmentation identification with no quantification, no detector modeling, and no recovery of sub-threshold signals; and top-down proteomics work quantifies intact proteoforms without any conservation constraint on the search space (Coupling CZE, Liquid-Phase Ion Mobility, to MS/MS for Quantitative Top-Down Proteomics). Neither is a competitor to the claim being made here.

Baselines

The naive reference point is standard peak picking with conventional intensity thresholds followed by ordinary database search on the untouched spectra — the "do nothing" pipeline that essentially all current proteomics runs. It is the correct floor because it defines the current effective noise threshold of the instrument; beating it means recovering and correctly quantifying fragment ions that this pipeline discards, measured as the lowest spike-in concentration at which recovered intensity remains linear (R² above 0.95) across serial dilutions.

The strongest existing competitor is the Pep2Prob transformer, which predicts the existence probability of each of 235 possible fragment ions per precursor and reports 95.3% existence accuracy with an L1 loss of 0.056 across 608,780 unique precursors drawn from 227 human HCD datasets (Pep2Prob Benchmark: Predicting Fragment Ion Probability for MS2-based Proteomics). It is the strongest because it is the best-characterized, benchmark-scale model of exactly the quantity we care about — whether a weak fragment ion truly exists — and its benchmark data will be reused directly. We will run it on identical spectral splits with an equal tuning budget (grid search over its hyperparameters matched to our calibration budget in GPU-hours), and specifically evaluate its accuracy restricted to the low-intensity fragment subset, which its headline 95.3% aggregates over and which we expect to be its weak regime. Beating it means higher existence accuracy on that low-intensity subset plus, crucially, calibrated intensities, which existence-probability prediction does not provide at all. The Dear-DIA XMBD autoencoder (Dear-DIA XMBD: Deep Autoencoder Enables Deconvolution of Data-Independent Acquisition Proteomics) is the second competitor for the chimeric-spectrum setting; no reported result found on Lacuna for Dear-DIA XMBD's quantitative performance, so we will run it ourselves on shared DIA data and compare false-fragment rates at matched recall.

The key ablation runs our full pipeline twice more: once with the calibrated response matrix but a uniform (conservation-free) initial estimate, and once with the conservation prior but no response matrix (a restricted database search over motif-containing peptides). The first isolates the conserved-prior ingredient, the second isolates the physics ingredient. The proposal stands only if the full combination beats both ablations on sub-threshold recovery; if the prior-free unfolding alone suffices, the conserved-targeting contribution is decorative and the claim must be narrowed.

Why not just…?

Why not just train a bigger deep model on more spectra? The evidence in the file says this hits diminishing returns at exactly the regime we target. AdaNovo already needed a 40% parameter increase (47.3M to 66.3M) and a dual-decoder mutual-information architecture just to tolerate noise and rare modifications (AdaNovo: Towards Robust De Novo Peptide Sequencing in Proteomics against Data Biases); the foundation model trained on 30 million spectra collapsed to 0.635 AUROC on phosphorylation when trained from scratch in a low-data regime versus 0.881 with pre-training (Foundation model for mass spectrometry proteomics), and its own multi-task fine-tuning degraded glycosylation performance while improving phosphorylation from 0.948 to 0.988 — evidence that scaling representations alone produces interference and fragility rather than robustness. More fundamentally, none of these models output physically calibrated intensities or transfer across collision regimes, a documented failure mode of instrument-dependent fragmentation libraries (PROSPECT). That said, this is a real threat: if a Pep2Prob-style transformer retrained on the low-intensity subset matches our recovery accuracy, the physics machinery is unnecessary — and the head-to-head low-intensity comparison in the experiments is designed precisely to settle that question rather than argue around it.

Why not just do targeted MRM for the clinical markers? Targeted multiple reaction monitoring is the clinical gold standard, and it works: the intracerebral hemorrhage cohort study quantified 22 serum biomarkers by MRM and achieved a multivariable AUC of 0.9524 for 90-day poor outcome, with low-abundance proteins like UCH-L1 (adjusted odds ratio 9.23) carrying the prognostic signal (Prognostic biomarkers of intracerebral hemorrhage identified using targeted proteomics and machine learning algorithms). But MRM is capped at a small predefined panel, must know the target sequence in advance, and is blind to divergent or mutated isolates — the same diagnostic-escape failure that motivates conserved-region design in the first place. Our method aims at the complementary regime: untargeted or DIA acquisition of many conserved targets at once, tolerant of sequence variation outside the invariant motif. We are explicit that MRM remains the accuracy ceiling, and the spike-in experiment uses MRM-measured reference concentrations as ground truth rather than as a defeated rival; if our unfolded intensities cannot track MRM references linearly at low concentrations, the proposal fails on its own terms.

Why not just use nucleic acid amplification on the conserved regions? Isothermal and qPCR assays targeting conserved genes are cheap, sensitive, and established (Rapid detection of influenza A viruses using a real-time reverse transcription recombinase-aided amplification assay). But they confirm only the presence of a nucleotide sequence and are blind to translation, post-translational modification, and protein-level host responses — a discordance documented directly in the antimicrobial-resistance surveillance literature, where genotype alone does not predict phenotypic resistance (Distribution and Antimicrobial Resistance Trends of Bloodstream Bacterial Isolates: A 10-year Single-Center Study in China). The proteomic readout is the point of this proposal, not an incidental choice, and the conserved-motif prior is what makes the proteomic readout robust to the same evolutionary drift that defeats hypervariable-target assays.

Proposed Experiments and Decisive Checks

The first experiment calibrates and validates the response matrix on PROSPECT, fitting the broadening and detection-efficiency terms on one subset of collision energies and validating on held-out energies; success is reconstruction deviation on known fragment intensities in the single-digit percent range, with the EJ-301 benchmark's 1.42% best agreement and sub-4.28% average (High-fidelity unfolding of 252Cf neutron spectra in EJ-301 scintillators) serving as the methodological precedent for what calibrated unfolding achieves in the analogue domain. The second experiment is the sub-threshold spike-in: synthetic peptides containing the conserved WFYDGYPT motif are serially diluted into a complex human background, acquired by DIA, and analyzed by all baselines plus both ablations; the decisive number is the limit of quantification where recovered intensity stays linear, compared against peak picking, Pep2Prob, and Dear-DIA XMBD under equal compute and tuning budgets. The third experiment tests cross-species robustness on the 9-species de novo sequencing benchmark of roughly 1.5 million spectra, where AdaNovo outperformed DeepNovo, PointNovo, and Casanovo on 8 of 9 species and improved PTM-bearing amino-acid identification by 3.6% to 9.4% (AdaNovo); our conservation prior should help most on the species where human-trained models transfer worst, so beating AdaNovo on its weakest species is the targeted result. The fourth experiment is clinical: reanalyze DIA-style acquisitions of the 22-biomarker intracerebral hemorrhage panel against the MRM reference values from the original cohort (Prognostic biomarkers of intracerebral hemorrhage), asking whether unfolded intensities for low-abundance markers like UCH-L1 and Serpin-A11 track the MRM measurements well enough to reproduce the panel's prognostic association. The decisive check for the whole program is the ablation contrast: if conserved-prior unfolding does not beat both prior-free unfolding and prior-only restricted search specifically at sub-threshold intensities, the central claim is falsified, and we would report which single ingredient — if either — carries the gain.

References

  1. High-fidelity unfolding of 252Cf neutron spectra in EJ-301 scintillators via calibrated response matrices and Monte Carlo simulations
  2. Discovery of a broad-spectrum monoclonal antibody recognizing a conserved, linear epitope WFYDGYPT on VP1 protein of Enterovirus A species
  3. Non-Human Primates Harbor Diverse Mammalian and Avian Astroviruses Including Those Associated with Human Infections
  4. AdaNovo: Towards Robust \emph{De Novo} Peptide Sequencing in Proteomics against Data Biases
  5. Dear-DIA XMBD : Deep Autoencoder Enables Deconvolution of Data-Independent Acquisition Proteomics
  6. Pep2Prob Benchmark: Predicting Fragment Ion Probability for MS2-based Proteomics
  7. Phonon-mediated Migdal effect in semiconductor detectors
  8. PROSPECT: Labeled Tandem Mass Spectrometry Dataset for Machine Learning in Proteomics
  9. A highly efficient indirect ELISA and monoclonal antibody established against African swine fever virus pK205R
  10. Identifying Influenza Viruses with Resequencing Microarrays
  11. Structural determination of Streptococcus pyogenes M1 protein interactions with human immunoglobulin G using integrative structural biology
  12. Coupling CZE, Liquid‐Phase Ion Mobility, to MS/MS for Quantitative Top‐Down Proteomics: Revealing Significant Proteoform Differences Between Healthy and Alzheimer's Disease Brains
  13. Foundation model for mass spectrometry proteomics
  14. Prognostic biomarkers of intracerebral hemorrhage identified using targeted proteomics and machine learning algorithms
  15. Rapid detection of influenza A viruses using a real-time reverse transcription recombinase-aided amplification assay
  16. Distribution and Antimicrobial Resistance Trends of Bloodstream Bacterial Isolates: A 10-year Single-Center Study in China