Public proteomics archives hold billions of raw tandem mass spectra. A tandem mass spectrum is the fragment "fingerprint" recorded when an instrument breaks a peptide apart. Many of these spectra come from peptides carrying a phosphate group, the chemical switch that kinases attach to proteins to turn signaling on and off. Almost none of these phosphopeptides were ever identified, because each lab searched its own files once, against its own narrow settings. The best learned detector of phosphorylated spectra was trained on only 30 million curated spectra moved to one place (Foundation model for mass spectrometry proteomics). That paper names the raw, uncurated remainder as untapped. Seismology shows how to reach such data without moving it. One group ran a small neural "phase picker" as about 145,000 container jobs directly against 1.3 PB of waveforms sitting in cloud storage. It finished in days for roughly $15,000 (A Global-scale Database of Seismic Phases from Cloud-based Picking at Petabyte Scale). This "compute-to-data" approach sends the code to the stored data instead of downloading the data.
We propose the same move for phosphorylation, with two changes that make it a research question rather than an engineering port. The first is a cascade. A very small, distilled detector screens every spectrum where it is stored, and the expensive models run only on what it flags. The second is an association step that links scattered per-spectrum calls into phosphosites with controlled error. Seismology's billions of picks left this linking step unsolved. The claim is specific and testable. A detector with a few million parameters, distilled from the foundation model, can keep nearly all of the teacher's phospho recall at a per-terabyte cost comparable to the seismic picker. Cross-run association then turns its noisy calls into a phosphosite atlas with lower false discovery than thresholding single spectra.
Core Claim
Three findings in the file motivate the design. First, model size sets the bill. AdaNovo needs about 7 hours on an A100 GPU for a single benchmark species (AdaNovo). NovoBench reports that such models run up to 4x slower than 5–8M-parameter predecessors (NovoBench). Running a 47–92M-parameter sequencer over billions of spectra therefore breaks the economics that made the seismic run feasible.
Second, the foundation model shows that a frozen encoder plus a small head separates phospho from non-phospho spectra well. It reports an AUROC (area under the ROC curve) of 0.948, and 0.948 is the number the cheap detector must approach. AUROC rises to 0.988 with multi-task training (Foundation model for mass spectrometry proteomics). So the detection signal exists and could, in principle, be distilled.
Third, a negative result constrains where the cheap filter may sit. NovoBench found that very clean spectra often lack key fragment ions, so precision first rises as noise rises (NovoBench). The natural cheap pipeline discards "low-quality" spectra first, using a quality score or AdaNovo-style down-weighting (AdaNovo). That ordering may silently throw away exactly the sparse spectra that pin down a phosphosite. Our design never applies a quality filter before the phospho detector. Whether this ordering matters is one of the experiments.
Research Program
The work has three parts.
Distillation. We distill the frozen Casanovo-based encoder plus its phospho head into a compact detector. The detector reads the precursor mass and the top-k peaks and outputs a phospho probability. We train it on MassIVE-KB labels and on teacher soft labels over a large unlabeled sample.
In-place cascade. We package the detector as a container that streams raw files from object storage and writes compact calls back. Each call records the spectrum ID, the probability, and the precursor. Only flagged spectra pass to a second stage. That stage runs the full foundation-model head, localizes the phosphosite against DeepPhospho-predicted fragment intensities (DeepPhospho), and, for spectra no database explains, runs a de novo sequencer.
Association. Flagged spectra are clustered across runs and datasets by precursor mass, retention behaviour, and second-stage peptide assignment. A phosphosite is reported only when independent runs support it. Error is controlled with a target–decoy scheme that places decoy phosphates on residues that cannot be phosphorylated.
The repository side is a real constraint. The seismic work benefited from archives already sitting in public cloud buckets, and it notes that data outside such programs stays disconnected (A Global-scale Database of Seismic Phases from Cloud-based Picking at Petabyte Scale). We will therefore pilot on whatever portion of MassIVE raw data can be read in place, either cloud-mirrored or on repository-hosted compute. We report results per terabyte rather than claiming archive coverage we have not achieved. The data-commons argument that each access barrier cuts participation sharply (Ten lessons for data sharing with a data commons) is why the output is a downloadable call table plus a reusable container. A bespoke portal would add exactly the kind of barrier that argument warns against.
Baselines
The do-nothing reference point is the status quo. Every public dataset keeps only the identifications its submitters produced. We measure how many atlas phosphosites appear in no submitter result on the processed slice.
The strongest per-spectrum competitor is the foundation model's phospho head itself (Foundation model for mass spectrometry proteomics). It reports phosphorylation AUROC 0.948 single-task and 0.988 multi-task. The cheap detector is evaluated on the same held-out MassIVE-KB phospho test split, with the same labels. "Beating" it is not the goal. Success means getting within 0.01 AUROC of 0.948 while costing at least 10x less per spectrum, measured on identical hardware. We will also run the teacher end-to-end with no cascade on a 1% archive slice and extrapolate its cost; this is the expensive upper bound on recall.
A cheap chemistry heuristic is the obvious non-learned competitor. Phosphoserine and phosphothreonine peptides often lose phosphoric acid (98 Da) on fragmentation, which leaves a recognizable peak. We implement a neutral-loss detector and give it the same tuning budget, meaning one threshold swept on the validation split. No reported result found on Lacuna for neutral-loss screening against learned detectors.
For de novo stages we use the NovoBench protocol. NovoBench gives no numeric precision values in the file. AdaNovo reports a 3.6–9.4% gain in PTM-residue identification on the Nine-species benchmark (AdaNovo).
Two ablations isolate the key ingredients. The first reverses the order and filters by spectrum quality before phospho detection; this tests the noise-paradox concern. The second removes association and reports sites from single-spectrum thresholds; this tests whether linking calls across runs is what controls error.
Why not just…?
Why not just download curated libraries and run the foundation model centrally, as its authors did? For labeled data this works, and it is how we train. The question is whether the uncurated remainder adds anything. A saturation curve will settle it: new phosphosites recovered as a function of the fraction of raw data processed. If the curve flattens after a small sample, then archive-scale in-place processing is unnecessary and we will say so. The deliverable then shrinks to the distilled detector and the association method.
Why not just reanalyze the pre-extracted abundance tables that repositories already host? Tables inherit the original search space. DeepPhospho shows that simply reanalyzing with a better library recovers up to 25.9% more phosphopeptides than the project's own library (DeepPhospho). Tables also carry abundance-dependent missingness that requires costly imputation (DreamAI). Spectra never searched for phosphorylation cannot appear in any table.
Why not just use the 98 Da neutral-loss rule as the screen? This is the most serious threat, because it is nearly free. It likely misses phosphotyrosine and spectra where the loss is weak, but that is an expectation, not a measured result. We compare recall stratified by residue type and by noise/missing-fragment bins. If the rule matches the distilled detector's recall, the cascade's first stage becomes the rule. The contribution then rests on association and the cost/recall audit.
Why not just buy faster networks? Optical-lightpath transfer such as Sector raises throughput but leaves every lab storing its own copy. The seismic comparison is days and $15,000 in place versus weeks to months of transfer. Our cost-per-terabyte measurement makes the same comparison for mass spectra.
Proposed Experiments
The decisive checks run in order, and each one can falsify the claim.
The first test is distillation fidelity. On the held-out MassIVE-KB phospho split, the compact detector must stay within 0.01 AUROC of 0.948 and reach at least 95% recall at its operating threshold.
The second test is the noise-paradox audit. Recall is binned by noise ratio and missing-fragment ratio, using the NovoBench definitions. We test whether quality-filter-first pipelines lose phospho recall disproportionately in the clean but fragment-poor bin. If they do not, the ordering constraint is unnecessary.
The third test is localization accuracy. On DeepPhospho's synthetic human and yeast phosphopeptides, where the true sites are known, we measure site-level accuracy of the second stage.
The fourth test is association. On the processed archive slice, site-level decoy false discovery with association must be clearly lower than single-spectrum thresholding at equal recall on the synthetic set.
The fifth test is economics. We measure wall-clock time and dollars per terabyte on preemptible instances. The result is reported against the teacher-only extrapolation and against the seismic run's roughly $11.5k per PB, computed from $15,000 over 1.3 PB. The bar is a figure of the same order.
References
- Foundation model for mass spectrometry proteomics
- A Global-scale Database of Seismic Phases from Cloud-based Picking at Petabyte Scale
- AdaNovo: Towards Robust \emph{De Novo} Peptide Sequencing in Proteomics against Data Biases
- NovoBench: Benchmarking Deep Learning-based \emph{De Novo} Sequencing Methods in Proteomics
- DeepPhospho accelerates DIA phosphoproteome profiling through in silico library generation
- Ten lessons for data sharing with a data commons
- DreamAI: algorithm for the imputation of proteomics data
- A peer-to-peer infrastructure for distributing large scientific data sets over wide area high-performance networks: experimental studies using wide area layer 2 services