Proteomics

Research questions

• Which proteins and proteoforms differ between conditions, and do they follow mRNA changes?
• Which signalling pathways are active (phosphoproteomics) and which kinases drive them?
• Which plasma proteins are biomarkers or drug targets, and which are causal (pQTL + Mendelian randomisation)?
• Which proteins form complexes, where are they located, and how do structures explain variant effects?
• Which drug binds which protein (chemoproteomics, thermal proteome profiling)?

How proteomic data are generated

1. SampleCells, tissue, plasma or serum (EDTA plasma preferred); avoid haemolysis and keratin contamination
2. Lysis & quantifyDetergent or urea buffers; protein amount by BCA assay
3. Reduce & alkylateBreak disulfide bonds (DTT/TCEP) and cap cysteines (iodoacetamide)
4. DigestTrypsin overnight (or S-Trap/SP3 workflows); optional TMT labelling or PTM enrichment
5. Clean upDesalt peptides (C18); measure peptide amount
6. LC-MS/MSNano- or micro-flow LC into an Orbitrap or timsTOF; DDA or DIA acquisition, 30–120 min per sample

Affinity platforms differ: Olink incubates plasma with pairs of DNA-tagged antibodies (a signal only forms when both bind the same protein), and SomaScan uses modified aptamers. Both convert protein amounts into nucleic-acid or fluorescence readouts and need only microlitres of plasma.

How bottom-up mass spectrometry works

1. Sample preparation: lyse, reduce and alkylate cysteines, digest with trypsin (cuts after K/R) into peptides of ~7–30 residues. Optional enrichment for PTMs (phospho: TiO₂, Fe-IMAC; ubiquitin: K-ε-GG antibody) or depletion of abundant plasma proteins.
2. Liquid chromatography separates peptides by hydrophobicity over 10–120 minutes; they are ionised by electrospray.
3. MS1 measures mass-to-charge (m/z) of intact peptide ions; MS2 fragments selected ions (CID/HCD, ETD) and measures fragment masses (b- and y-ions), which encode the sequence.
4. Acquisition: DDA (data-dependent; picks the top-N most intense ions, giving stochastic missing values), DIA (data-independent; fragments all ions in m/z windows, more complete and reproducible, more complex spectra), targeted (SRM/PRM; predefined peptides, best for validation).
5. Identification: match spectra against theoretical spectra from a protein database (MSFragger, Comet, Andromeda, Sage, Mascot) or spectral libraries (often predicted by deep learning: Prosit, AlphaPeptDeep, MS²PIP). Control FDR with target–decoy competition at peptide-spectrum match, peptide and protein levels (typically 1%); rescoring with Percolator or MS²Rescore.
6. Protein inference: peptides shared between proteins create ambiguity, resolved into protein groups by parsimony.
7. Quantification: label-free (MS1 intensity, MaxLFQ), isobaric tags (TMT/TMTpro: up to 18 samples multiplexed, measured at reporter ions; ratio compression from co-isolation), SILAC (metabolic labelling), or DIA fragment intensities.

Statistical analysis

MS proteomics data have features that RNA-seq tools do not handle:

• Missing values that are not random: low-abundance peptides are missing because they fell below detection (MNAR), while others are missing by chance (MCAR). Imputation choice changes results (minimum-value, left-censored QRILC or MinProb for MNAR; kNN or random forest for MCAR). Models that handle missingness directly (MSstats, proDA, limpa) are preferable to blind imputation.
• Log-normal intensities: analyse log₂ intensities with linear models (limma, MSstats, MSstatsTMT, DEqMS which accounts for the number of peptides per protein).
• Normalisation: median centring, variance-stabilising normalisation (vsn), TMT reference channels (IRS) for multi-batch designs.
• PTM site localisation probabilities (ptmRS, PTMProphet), and normalising PTM changes to protein abundance.
• Kinase activity inference from phosphosites: KSEA, PTM-SEA, the Kinase Library (Johnson et al. 2023).

Structure and interactions

AlphaFold2 (2021) predicts single-chain structures with near-experimental accuracy for many proteins; the AlphaFold Protein Structure Database holds over 200 million predictions. AlphaFold3 (2024) and open alternatives (Boltz-1/2, Chai-1, RoseTTAFold All-Atom) predict complexes with nucleic acids, ligands and modifications. ESMFold predicts structure from a protein language model without multiple-sequence alignments. pLDDT and PAE confidence scores must be read before trusting a model: low pLDDT often marks intrinsically disordered regions. Interaction resources include STRING (integrated scores), BioGRID and IntAct (curated), BioPlex and HuRI (systematic AP-MS and Y2H maps), and CORUM (complexes).

Tools

Benchmarks