Computational Proteomics

The translational proteomics working group defined translational proteomics as a stepwise process linking basic proteomics research to clinical and population-level impact. This includes:

• $\mathrm{■}$

  Experiments in model systems (cell lines, mouse models).

• $\mathrm{■}$

  Studies on human biospecimens (tissues, fluids).

• $\mathrm{■}$

  Clinical application (diagnostics, patient stratification, therapy guidance).

• $\mathrm{■}$

  Integration into routine decision-making (e.g. tumor boards).

• $\mathrm{■}$

  Extension to population health and precision medicine at scale.

The group agreed that translational proteomics is not only about “measuring patient samples,” but about structuring and standardizing the entire value chain: sample acquisition, metadata capture, MS acquisition, data processing, integration with other omics, interpretation, and delivery of clinically actionable output.

Data quality and interoperability: Most proteomics datasets are not captured with standardized metadata, ontologies, or controlled vocabularies suitable for clinical use. Heterogeneous assay formats, missing metadata, and inconsistent annotation impede reuse and prevent robust comparisons across sites.

Underpowered cohorts: Many studies operate at N too small for reliable biomarker development, particularly in rare diseases. Cohort fragmentation across institutions and7 slow/complex material transfer agreements impede the assembly of sufficiently powered datasets.

Workflow robustness and reproducibility: Clinical deployment requires durable, validated, auditable workflows. Current research pipelines are often bespoke, sensitive to batch effects, and not easily transferable to regulated environments.

Motivation and incentives: Clinically relevant work often requires repetitive, “boring” targeted assays and rigorous QC. Such work is rarely rewarded academically and is often unfunded. As a result, essential confirmatory and longitudinal measurements are underperformed.

Interpretability and clinical usability: Even when high-quality measurements exist (e.g. phosphoproteomics, glycoproteomics, immunopeptidomics), delivering a clear, defensible recommendation to a tumor board remains difficult. Clinicians want interpretable, validated evidence (what pathway is active? which drug class is likely to work?), not just feature lists. Regulatory and legal constraints: Clinical proteomics data raise privacy, compliance, and regulatory concerns analogous to those in genomics. Institutions are cautious about releasing raw data, which complicates model development and validation.

Standardized ontologies and metadata schemas (including disease context, acquisition parameters, and sample processing information) are required to make datasets reusable across institutions. Scalable, transparent pipelines are needed to analyze imperfect, real-world study designs (non-ideal controls, heterogeneous sampling times, varying platforms). Multi-omics integration must move beyond naive overlap of “differential features” and instead support mechanistic interpretation and phenotype prediction. Cloud-based or federated frameworks may be required to enable collaborative analysis when data cannot leave clinical boundaries. Quantitative modeling should address proteoforms, phosphorylation, glycosylation, immunopeptidomics, etc., rather than collapsing information to a single “protein ID.”

To address the chronic shortage of large, harmonized, clinically relevant datasets, the group proposed ENIGMA: an international, multi-site initiative to create an AI-ready proteomics resource on the order of >100,000 samples.
The initiative is structured in phases:
Phase 1 (Pilot): A few thousand well-controlled samples (initially including defined mouse lines and select human material) are processed to establish baseline protocols, metadata standards, benchmarking pipelines, and QC criteria.
Phase 2 (Scale-up): Tens of thousands of samples are acquired across multiple mass spectrometry laboratories worldwide under harmonized DIA-style acquisition. Heterogeneity (instrument type, sample preparation differences) is explicitly captured rather than avoided, to allow development of normalization, batch correction, and cross-lab integration strategies.
Phase 3 (Translation): The trained models and pipelines are applied to human cohorts, including rare disease and clinically complex cases, with the goal of informing stratification, therapeutic targeting, and mechanism-of-action hypotheses.
Critical design features:

• $\mathrm{■}$

  Use of leftover tissues and archived material to reduce ethical/consent burden and accelerate scale.

• $\mathrm{■}$

  Systematic metadata capture (including tissue, condition, preparation, instrument, acquisition settings), with machine learning support for harmonization and outlier detection.

• $\mathrm{■}$

  Community governance around data ownership, licensing, authorship, and credit.

• $\mathrm{■}$

  Defined benchmark subsets for public release, similar in spirit to CPTAC and AlphaFold-style “challenge” datasets.

The working group concluded that translational proteomics requires a coordinated village: clinicians to articulate clinical questions; experimentalists to generate high-quality material; computational groups to build robust, interpretable pipelines; and funding/industry partners to sustain longitudinal measurement. The ENIGMA plan will serve both as a rallying point for funding proposals and as an anchor for future community standards in translational proteomics.

Federated learning (FL) is widely discussed in biomedical AI but is not yet common in proteomics practice. The group identified why: For most discovery proteomics, data are still deposited centrally (e.g. PRIDE, MassIVE). Central pooling is simpler than setting up FL. FL infrastructure is non-trivial: institutions must maintain local compute, synchronize software/hardware, address batch effects and fairness in distributed model updating, and handle versioning and auditing. Repositories themselves would bear significant cost to orchestrate FL at scale. However, participants agreed that this will change as proteomics becomes clinically embedded. Clinical and translational datasets (especially targeted assays, patient-derived longitudinal data, immunopeptidomics, and phospho-signaling panels) are increasingly held locally for regulatory and privacy reasons. In that emerging setting, FL – or at least controlled-access, privacy-aware model training – becomes essential.

The group emphasized that the blockers are not just technical but social:

• $\mathrm{■}$

  Privacy/compliance anxiety: When rules are unclear, many investigators choose not to share data at all.

• $\mathrm{■}$

  Fear of being scooped: High-value clinical datasets are expensive to generate, and groups are reluctant to expose them prior to publication without formal recognition.

• $\mathrm{■}$

  Fear of policing: Investigators worry about reputational risk if preliminary or messy data are scrutinized out of context.

Participants noted that vast quantities of clinically interesting proteomics data remain unpublished and effectively invisible. Releasing even partial access to these datasets (or enabling them to participate in FL-like analysis) would massively accelerate method development and translational validation.
Proposed actions:

• $\mathrm{■}$

  Draft a community position / opinion piece that (a) documents current barriers to sharing unpublished proteomics data, (b) argues for explicit dataset credit and citation, (c) proposes governance and recognition mechanisms for dataset contributors, and (d) frames FL as one of several approaches (not the only one) for responsibly leveraging sensitive data.

• $\mathrm{■}$

  Promote dataset-level DOIs, ORCID linkage, citation tracking, and usage metrics as first-class research outputs.

• $\mathrm{■}$

  Encourage journals and funders to recognize data notes / dataset publications and not penalize manuscripts that build on previously described datasets.

The group critically assessed the state of “multi-omics integration” in proteomics-driven biology and translation. The conclusion was that true integration is still rare. Core challenges:

• $\mathrm{■}$

  Different omics layers (genomics, transcriptomics, proteomics, phosphoproteomics, glycoproteomics, metabolomics, lipidomics, immunopeptidomics, spatial MS imaging) are often collected on different material, at different time points, using incompatible extraction chemistries.

• $\mathrm{■}$

  Biological timescales differ: DNA is static, RNA is dynamic on hours, protein abundance and PTMs reflect turnover and signaling, metabolites respond on minutes. Naively correlating steady-state measurements across these layers can be misleading.

• $\mathrm{■}$

  Statistical control is underdeveloped. Integrating multiple high-dimensional datasets multiplies the number of hypotheses tested, but FDR control is often applied separately per layer, if at all.

• $\mathrm{■}$

  QC practices are inconsistent. Many studies still lack standardized spike-ins, technical controls, and batch assessment across all omics layers.

The group distinguished three levels of integration:

• $\mathrm{■}$

  Late integration: Interpret each omics layer independently, then compare or combine conclusions (e.g. pathway enrichment or overlapping differentially regulated features). This is the most common today.

• $\mathrm{■}$

  Mid-level / latent integration: Learn joint low-dimensional representations or network structures across layers (e.g. linking phosphoproteomics to transcriptomics through signaling pathway models).

• $\mathrm{■}$

  Early fusion: Combine raw or minimally processed quantitative data across layers into a single model. This is the most ambitious but also most fragile with respect to batch effects, missingness, and sampling asynchrony.

Use cases discussed:

• $\mathrm{■}$

  Proteogenomics to detect neoantigens, fusion proteins, and tumor-specific sequence variants for immunopeptidomics.

• $\mathrm{■}$

  Phosphoproteomics + total proteomics to distinguish signaling changes from protein abundance changes.

• $\mathrm{■}$

  Glycoproteomics to refine clinically relevant biomarkers (e.g. glycoform-specific PSA outperforms total PSA).

• $\mathrm{■}$

  Spatial MS (laser capture microdissection, MSI-guided microproteomics/metabolomics/glycomics) for tumor microenvironment profiling.

Actionable needs:

• $\mathrm{■}$

  Clear guidance on experimental design for clinically realistic studies, acknowledging that perfect co-sampling is often impossible in the hospital setting.

• $\mathrm{■}$

  Shared QC expectations across layers (technical controls, spike-ins).

• $\mathrm{■}$

  Agreement on how to claim “integration”: studies should report whether they did late, mid-level, or early fusion, rather than labeling any multi-assay project as “multi-omics integration.”

• $\mathrm{■}$

  Availability of paired, public benchmark datasets suitable for method development.

The cross-over discussions also addressed quantification and statistical analysis for glycomics and glycoproteomics, linking glyco specialists, statisticians, and ML practitioners.
Key points:

• $\mathrm{■}$

  Glycopeptide intensities are not directly comparable across glycoforms, because different glycans alter ionization and fragmentation efficiency.

• $\mathrm{■}$

  Missing values should not always be blindly imputed. If an entire glycoform is absent in one condition, treating that as “low abundance” can generate false positives.

• $\mathrm{■}$

  Normalization strategies must reflect biology. Global scaling can erase true global glycosylation shifts; site-focused normalization may be more appropriate.

• $\mathrm{■}$

  Site occupancy and glycoform distribution are biologically meaningful readouts but require modeling analogous to PTM-centric statistical frameworks.

Planned work:

• $\mathrm{■}$

  Evaluate MSstats/MSstats-PTM–style approaches where glycoforms are treated as modified analytes, using controlled datasets (e.g. exoglycosidase-treated samples, known congenital glycosylation defects) to benchmark differential analysis, normalization, and missing value handling.

• $\mathrm{■}$

  Define minimal reporting requirements for glyco-quantitative studies, to make them interpretable and reusable in translational pipelines.

Glycomics and glycoproteomics remain computationally and analytically challenging due to structural complexity (branching, linkage, isomerism), diverse acquisition strategies, non-uniform software support, and historically limited standardization. The glyco working group concentrated on two urgent needs: (i) harmonizing identification and reporting so that results are comparable across software and labs, and (ii) establishing reliable quantitative and statistical practices, especially for glycopeptide-level measurements.

A core outcome was agreement that glycan and glycopeptide search results must become machine-readable, comparable, and traceable across tools. The group recommends:

• $\mathrm{■}$

  Use of GlyTouCan IDs in all reported results, at minimum at the composition level, with increasing specificity (topology, linkage) where supported by evidence.

• $\mathrm{■}$

  Explicit declaration of structural specificity: Results should indicate whether the assignment is composition-only, topology (branching pattern without full linkage certainty), or fully linkage-resolved.

• $\mathrm{■}$

  Subsumption / hierarchical comparison: Since different tools resolve glycans to different specificity levels, there must be a way to “collapse” structures to a shared lower-specificity representation for comparison and benchmarking.

• $\mathrm{■}$

  Shared glycan reference sets: The group began assembling a curated “human adult reference glycan list,” analogous to a reviewed FASTA for proteins. Each participating lab/software group will contribute the glycan lists they currently search (e.g. as used in Byonic, MSFragger-glyco, pGlyco, Protein Prospector, GlycanFinder, etc.). These will be merged, deduplicated, and annotated with GlyTouCan ID, species/context, linkage class (N- vs O-linked), biological source (plasma, tissue, cell line), and whether the glycan is free or peptide-conjugated.

• $\mathrm{■}$

  The aim is to generate a community-supported glycan search space that is biologically grounded, not “everything in GlyTouCan,” and thus suitable for routine glycoproteomics searches and cross-tool benchmarking.

The group addressed statistical confidence, which currently lags behind peptide-centric proteomics:

• $\mathrm{■}$

  Glycan/glycopeptide identification often lacks robust FDR control, especially at the level of glycan topology or isomer resolution.

• $\mathrm{■}$

  Scoring schemes must consider diagnostic fragment ions, instrument- and method-specific fragmentation behavior (e.g. stepped HCD vs ETD/EThcD), and the ability to distinguish core vs antenna fucosylation, high-mannose vs complex glycans, etc.

• $\mathrm{■}$

  Structural-level FDR (e.g. topology-level “localization” of glycan features on the peptide) was viewed as analogous to PTM localization scoring in phosphoproteomics.

Participants presented ongoing work on topology-aware scoring frameworks that progressively reduce candidate lists using diagnostic ions and instrument-aware fragmentation rules, then apply multi-level FDR (PSM-level, composition-level, topology-level). The consensus was pragmatic: “some FDR is better than none,” provided the confidence level is clearly stated. Work remains to prevent overconfidence in cases where only one plausible glycan remains in the database but the MS/MS evidence is insufficient to distinguish isomers.

A second major focus was quantification. Glycopeptide signals are strongly influenced by the attached glycan, so intensity is not a straightforward proxy for occupancy or abundance. The group discussed:

• $\mathrm{■}$

  Handling missing values: imputing partially missing precursors may be acceptable, but imputing entire missing glycoforms is risky and can create false biological conclusions.

• $\mathrm{■}$

  Normalization: global median-centering or total-signal normalization can be misleading if overall glycosylation shifts biologically. Alternative strategies include normalizing within a site to the dominant glycoform, or modeling site occupancy relative to the unmodified protein level.

• $\mathrm{■}$

  Statistical modeling: Treating glycoforms as site-specific PTMs suggests that tools such as MSstats/MSstats-PTM could be adapted by treating each glycoform as the “modified species.”

The group proposed benchmarking known perturbation datasets (e.g. exoglycosidase-treated samples, congenital disorders of glycosylation, cell-surface enrichment studies) to evaluate differential analysis, missing value handling, and normalization strategies. This work connects directly to translational aims (e.g. leveraging glycosylation patterns in diagnostics and tumor board discussions) and to ML aims (e.g. training glyco-aware spectral/RT predictors, or “glyco-Prosit” models).

The working group advanced two manuscript efforts that originated in a previous Dagstuhl meeting:

• $\mathrm{■}$

  Best-practices/tutorial manuscript

• $\mathrm{■}$

  Introduces newcomers to protein glycosylation analysis, including released glycans and intact glycopeptides.

• $\mathrm{■}$

  Defines key terminology (composition, topology, linkage-defined structure).

• $\mathrm{■}$

  Recommends that reported results always include GlyTouCan IDs and explicitly state confidence and specificity level.

• $\mathrm{■}$

  Summarizes common pitfalls in identification, quantification, and interpretation.

• $\mathrm{■}$

  Aligns with MIRAGE-style reporting expectations.

During the seminar, section leads were assigned, timelines were agreed upon, prior contributors were re-engaged, and a coordinated writing sprint produced 21 pages of draft text.

The glycomics/glycoproteomics group delivered concrete progress toward harmonized reporting, FDR-aware scoring, and quantitative interpretation, and established a clear writing and dissemination plan. The group also linked their agenda to ML (glyco-aware prediction models, benchmark datasets) and translational proteomics (inclusion of glycan biology in clinically oriented pipelines).

The ML working group took a critical view: the field currently rewards incremental improvements (e.g. 2% gain in spectral prediction accuracy) without necessarily delivering interpretability, robustness, reproducibility, or biological/clinical insight. The group argued for a rebalancing of values toward sustainability and impact.

Four core themes structured the discussions:
(i) Software Quality, Standards, and Publication Expectations
Participants assessed the adequacy of existing frameworks such as the DOME recommendations for reporting ML in proteomics/metabolomics and FAIR4RS for FAIR research software. They concluded that these frameworks are necessary but incomplete. Key gaps:

• $\mathrm{■}$

  Maintainability and testability: ML code should be modular, documented, versioned, licensed, and accompanied by retraining instructions and tests. These expectations are rarely enforced.

• $\mathrm{■}$

  Adherence to community standards: ML tools must input and output established HUPO-PSI formats (mzML, mzIdentML, mzTab, SDRF, ProForma, etc.) where applicable, to ensure interoperability.

• $\mathrm{■}$

  Bias and applicability domain: Authors should explicitly analyze dataset composition (organism, tissue, modification class, charge states, instrument type), identify biases (e.g. human and HLA bias in immunopeptidomics, charge state bias, peptide length bias), and articulate where the model should and should not be used.

• $\mathrm{■}$

  Novelty and publication worthiness: DOME specifies how to describe an ML method, but not whether the method represents a meaningful advance. The group argued that journals and reviewers should demand either conceptual novelty, new interpretability, improved robustness/maintainability, or concrete biological/clinical utility – not just a small numeric gain.

Outcome: The group began drafting a manuscript that extends previous Dagstuhl-driven work (including “Interpretation of the DOME Recommendations for Machine Learning in Proteomics and Metabolomics” and recent guidance on FAIR/open-source research software in proteomics) by incorporating the EVERSE research software quality dimensions (modularity, reusability, analysability, modifiability, testability, sustainability). The goal is a reviewer/author checklist that goes beyond reporting and into quality, longevity, and responsible reuse.
(ii) Training Data, Metadata, Uncertainty, and Benchmarking
Robust ML depends on training data that is complete, well-annotated, and accompanied by uncertainty estimates. The group emphasized that current proteomics and glycoproteomics datasets often lack even basic metadata (sample type, organism, sex, preparation method, instrument settings), which blocks fair benchmarking and leads to hidden biases. Priorities:

• $\mathrm{■}$

  Centralized metadata capture: Multiple ongoing efforts (curated SDRF files, reprocessing pipelines, LLM-based annotation of manuscripts and raw mzML headers, large-scale reanalyses such as MassIVE, PRIDE, PeptideAtlas, and MassNet) should not remain siloed. A shared portal under an existing infrastructure (e.g. ProteomeXchange) could collect experimental design metadata, derived identifications/quantifications, QC metrics, and completeness scores.

• $\mathrm{■}$

  Gold standard datasets: For each ML task (fragmentation prediction, retention time prediction, identification, quantification, modification localization, glycopeptide assignment, etc.), the community should define benchmark datasets that include raw data, identifications, quantitative outputs, and experimental design metadata in standard formats.

• $\mathrm{■}$

  These benchmarks must reflect biological and technical diversity (different tissues, instruments, acquisition modes) rather than a single “easy” mixture.

• $\mathrm{■}$

  Uncertainty propagation: Both PSM-level confidence and metadata confidence should be carried through into training. Models should not be trained exclusively on idealized “perfect” identifications; probabilistic or Bayesian treatment of noisy/ambiguous cases is encouraged.

• $\mathrm{■}$

  Synthetic data: Simulated/synthetic data can support benchmarking, experimental design, rare-event modeling, and privacy-preserving analysis. But the group raised integrity and governance concerns: synthetic vendor-format raw files must be clearly watermarked and traceable, to prevent fraud and to allow repositories and journals to distinguish simulated from measured data. Synthetic data are not a substitute for high-quality real data in final model evaluation.

(iii) Interpretability and Explainable AI (XAI)
The group emphasized that interpretable ML is essential if models are to influence biological discovery, clinical triage, or mechanistic reasoning. Two levels were distinguished:

• $\mathrm{■}$

  Explainability: Tools such as SHAP values, saliency maps, causal/graphical models, and feature attribution methods can reveal which features drive predictions in a given model or for a given sample.

• $\mathrm{■}$

  Interpretability: The biological and biochemical meaning of those features must then be contextualized. For example, a classifier that distinguishes tissue-of-origin should identify pathways and protein/glycoform signatures plausibly linked to that tissue, rather than artifacts such as systematic missingness or batch effects.

Planned output: The group outlined a manuscript arguing for explainable and interpretable AI in MS-based proteomics and glycoproteomics. This manuscript will catalog use cases (e.g. tissue-of-origin prediction, phospho-signaling interpretation, glycoform-driven biomarker hypotheses, failure analysis in de novo sequencing), pitfalls (spurious correlations, hidden batch effects), and recommendations for reporting.
(iv) Education and Community Infrastructure
Interest in ML for proteomics and glycomics is accelerating, but training materials remain fragmented. Rather than generating entirely new courses, the group agreed to curate and maintain a community-driven “Awesome Proteomics/ML” style resource, linked to ProteomicsML and ELIXIR TeSS. This resource will collect:

• $\mathrm{■}$

  Introductory material for ML researchers entering proteomics (instrument basics, data structure, pitfalls).

• $\mathrm{■}$

  Practical material for experimentalists entering ML (basic statistics/ML literacy, bias analysis, model evaluation).

• $\mathrm{■}$

  Glyco-specific onboarding resources, which are currently underrepresented compared to proteomics.

The group discussed adding vetted content to TeSS and ProteomicsML, tagging material by audience and task, and exploring lightweight retrieval-augmented assistants based on this curated corpus.

The ML working group shifted attention from “How do we get slightly better predictions?” to “How do we build models, datasets, software, and training practices that the community can trust, reuse, and interpret – and that actually matter biologically and clinically?” Deliverables in progress include:

• $\mathrm{■}$

  A standards/checklist manuscript integrating DOME, FAIR4RS, and EVERSE for ML in proteomics and glycomics.

• $\mathrm{■}$

  A perspective on explainable and interpretable AI in MS-based proteomics.

• $\mathrm{■}$

  A shared educational and onboarding resource, to be maintained in community infrastructure.