22 proteomics process pack

MethylPipeline supports proteomics as a third omics modality alongside DNA methylation and RNA-Seq. The execution platform (DomainProgram, typed actions, scheduler, CAAS, cfg/wf) is shared; the front of the pipeline (mass-spec search or panel ingest) is new and the back reuses the same samples x features seam as RNA-Seq. This chapter is the operator guide.

For the full topology and diagrams see the proteomics end-to-end workflow.

Modality selection¶

Set regulatory.primary_modality: proteomics (or methyl-study-init --modality proteomics). Modality is the assay axis and is independent of primary_analyte.

Ingest modes¶

actionConfig.proteomics_quant.ingest_mode picks the path (like RNA quant_mode):

`ingest_mode`	Action	GPU	Output
`dia` (default)	`sample.diann` (+ optional `sample.dl_rescore`) → `sample.register_abundance`	Yes (GH200)	DIA-NN report → `abundance.h5`
`dda`	`sample.sage` → `sample.register_abundance`	No (CPU)	Sage search + LFQ → `abundance.h5`
`panel`	`sample.ingest_panel`	No (CPU)	Olink NPX / SomaScan RFU / open matrix → `abundance.h5`

Set rescore: true under proteomics_quant to add Prosit deep-learning rescoring on the DIA path. Casanovo de novo (sample.casanovo) is available as a complementary GPU action.

DDA via Sage. ingest_mode: dda runs Sage (Apache-2.0, Rust) — a fast, open database-search engine that replaces the license-encumbered MSFragger/FragPipe. It is CPU-only and multi-arch, so it lands on cheap CPU workers (not the GH200s). Sage knobs live under actionConfig.proteomics_quant (missed_cleavages, precursor_tol_ppm, fragment_tol_ppm, variable_mods — e.g. phospho on S/T/Y for tau phosphopeptide discovery, lfq). Set METHYL_SAGE_IMAGE (or install the sage binary). Sage produces protein LFQ intensities normalized into the same abundance.h5 contract as DIA-NN and panels.

Reference assets (site)¶

Pin proteomics references once per cluster under site proteomics_reference:

"proteomics_reference": {
  "protein_fasta": "/work/genomes/proteomics/GRCh38/uniprot/UP000005640_9606.fasta",
  "spectral_library": "/work/genomes/proteomics/GRCh38/diann/human_plasma.speclib",
  "prosit_model": "/work/genomes/proteomics/models/prosit/intensity_transformer",
  "casanovo_model": "/work/genomes/proteomics/models/casanovo/casanovo_massivekb.ckpt"
}

DIA-NN can run library-free with just protein_fasta. DL model weights should be pinned as cfg reference_assets for reproducibility.

GPU on GH200¶

sample.diann / sample.dl_rescore / sample.casanovo run on the same Lambda/Nebius GH200 GPU VMs as Parabricks, each behind its own image env (METHYL_DIANN_IMAGE, METHYL_PROSIT_IMAGE, METHYL_CASANOVO_IMAGE) and gated by GPU availability. On ARM64 (Grace) the images must be linux/arm64 or multi-arch — see the GPU worker runbook. Panel ingest is CPU-only and lands on non-GPU workers.

Downstream¶

pipeline.protein_de_select builds the cohort samples x proteins matrix (log2 + median-normalize + per-protein min imputation), ranks by Welch differential abundance, trains the shared tabular classifier (optionally stacking covariates), and writes a recurrence-friendly protein panel for validation.stability. feature_mode is proteomics_abundance.

Example run¶

methyl-workflow-run \
  --program workflow_engine/domain/fixtures/sample_prep_proteomics.program.json \
  --context '{"projectPath":"/work/projects/my-prot-study/configs/project_Healthy_vs_Disease.json","pipelineProfile":"proteomics_research"}'

Not included¶

Sage results → spectral library → DIA-NN (DDA-assisted DIA) — a later optional extension.
TMT/iTRAQ labeled quant (LFQ only this iteration).
MSFragger/FragPipe is intentionally not used (commercial license); Sage is the open DDA engine.