flowchart TD A["INS gene (DNA)"] -->|"transcription"| B["INS mRNA"] B -->|"translation"| C["preproinsulin"] C -->|"signal peptide removed"| D["proinsulin"] D -->|"folds, forms disulfide bonds,<br/>C-peptide cut out"| E["insulin + C-peptide"]
Lecture 8 — Proteomics
Why proteins · sample types · the plasma proteome · measurement technologies · NPX · QC and batch effects · PWAS · pQTLs · protein MR · colocalization · from association to mechanism · reading: eight platforms compared
DNA is fixed for life, but proteins change with diet, disease, drugs and age, and they are what actually does the work in cells. Most drugs act on proteins. That makes the proteome, all the proteins in a sample, an attractive layer for finding biomarkers and drug targets.
Measuring proteins is harder than reading DNA. Plasma contains proteins at concentrations that differ by many orders of magnitude, the same gene can give several protein forms, and most methods measure how much protein there is, not how well it works. Modern platforms (Olink, SomaScan) now measure thousands of proteins in tens of thousands of people, which allows proteome-wide association studies (PWAS), the protein version of a GWAS.
The second half connects proteins back to genetics. A SNP that changes a protein’s level is a pQTL. If it sits next to the protein’s own gene (cis-pQTL), it is a good instrument for protein Mendelian randomization: does genetically higher protein cause disease? The lecture closes the course’s PCSK9 story end to end.
This page covers why and where to measure proteins, the three technology families, what the numbers mean (NPX), QC and batch effects, PWAS, pQTLs (cis vs trans), protein MR and its specific pitfalls, colocalization and the path from association to mechanism.
NoteSources and session
Slides: “Lecture 8 proteomics.pdf”, delivered in two parts.
Reading material (added 2 October; marked (reading) in the text): Kirsher et al. (2025), “Current landscape of plasma proteomics from technical innovations to biological insights and biomarker discovery”, Communications Chemistry 8:279. Worked through in 8.13.
Rough plan: Week 4, day 1. High-dimensional data, proteomics.
Part 1: Proteomics fundamentals
8.1 Why study the proteome?
Contrast from the molecular-layer table (Lecture 1): DNA is fixed, while RNA and proteins are dynamic, affected by diet, disease, drugs, environment and age. Proteins matter because they carry out cellular functions, act as signalling molecules, respond to disease/environment/diet/drugs/age, and are the targets of most drugs (Leqvio and Repatha in Lecture 3).
Worked example: insulin biosynthesis. “One gene” does not always mean “one direct protein”:
(Insulin is a signalling molecule central to metabolism, obesity and diabetes; compare the glucose/insulin example in Lecture 1.)
Proteins are not 1-to-1 with genes. Processes that multiply protein products:
| Process | Example |
|---|---|
| Alternative splicing | One gene’s transcript spliced into several mRNA/protein variants |
| Cleavage | Insulin from proinsulin |
| Post-translational modification (PTM) | Phosphorylation, glycosylation, cleavage |
| Different cellular localization | Same sequence behaves differently depending on where it ends up |
| Protein complexes | Proteins assemble into larger units (quaternary structure, Lecture 1) |
Net effect: the proteome is more dynamic and complex than the genome.
ImportantProtein amount ≠ protein function
More protein does not automatically mean more function (though in the simplest case it does). The same amount can have different function, e.g. after a GoF or LoF mutation (the PCSK9 examples in Lecture 3): a functionally altered protein can be present at a normal concentration. Most proteomics methods measure amount (abundance), not function.
Every gene is translated into exactly one protein product
Splicing, cleavage and modifications create extra variants
Proteins are copied directly from other proteins in the cell
The genome changes continuously with diet and age
Alternative splicing, cleavage, post-translational modifications, localization and complex formation all multiply the number of distinct protein products per gene.
8.2 Where can proteins be measured?
| Sample type | Trade-off |
|---|---|
| Plasma / serum | Easily accessible; systemic, whole-body signal |
| Tissue | Needs a biopsy, but closer to the site of biology |
| Cells | Cell-type-specific mechanisms |
| Urine | Reflects kidney/urinary tract function and filtering |
| CSF (cerebrospinal fluid) | Relevant for neurological disease, but hard to obtain |
| Saliva, etc. | Accessible but more local; generally lower complexity/information than blood |
8.3 The plasma proteome
Many organs, including blood cells, release proteins into plasma. A protein’s circulating level is the net balance of production/synthesis, release, distribution, degradation, uptake and clearance, so a change in plasma level can reflect a change in any of these, not necessarily in how much is made. Plasma integrates signals from many tissues: some proteins come from one tissue and reflect it specifically; for others, the tissue you care about can drown in signal from other tissues.
Why plasma is attractive for diagnostics: easy to access and informative, a mixture carrying signal from the liver, immune cells, adipose tissue, muscle, damaged tissues (injury releases proteins, see troponin below) and secreted hormones/signalling proteins.
Measurement challenges in plasma
- some proteins have low abundance and/or are technically hard to measure;
- isoforms complicate quantification;
- signalling proteins and cytokines are often at very low concentration;
- mass vs count units (mg/dL vs mmol/L) can disagree because isoforms have different molecular weights;
- some proteins (e.g. albumin) are hugely abundant: plasma proteins span an enormous dynamic range (many orders of magnitude), a major challenge for any multi-protein assay.
Worked example: troponin. Normally very low in plasma. After myocardial injury: cardiomyocyte injury → troponin released from damaged cells → plasma troponin rises. Unlike a constitutively secreted protein, troponin in blood signals damage, not normal secretion (the “disease/tissue injury” biomarker type from Lecture 1).
8.4 Three technology families
| 1. Antibody-based | 2. Mass spectrometry | 3. High-throughput affinity | |
|---|---|---|---|
| Examples | ELISA, clinical immunoassays, Western blot | LC-MS | Olink, SomaScan |
| How | An antibody binds its target protein, giving a signal proportional to the amount | Protein digested into peptides with trypsin (cuts at lysine and arginine) → separated by liquid chromatography (LC) → charged by electrospray ionization → masses measured (“mass spectrum 1” = intact precursor peptide masses) | Each protein binds a specific engineered affinity reagent; the signal estimates relative abundance |
| Scope | Highly targeted: one or a few proteins | Can do discovery (unknown proteins/changes) | Thousands of proteins per sample |
| Quantification | Quantitative with calibration standards | Often calibrated with labelled control peptides | Relative abundance, not absolute |
| Strengths | Basis of routine clinical chemistry | Detects isoforms; discovery | Designed for large cohorts, biomarker discovery, PWAS |
| Weaknesses | Few proteins at a time | Technically complex; dynamic range needs dilution/preparation | Relative values only; assay binding can be affected by variants (8.11) |
Real-world example: UK Biobank (UKBB) proteomics: about 50,000 individuals, about 3,000 proteins each.
ELISA, since it is quantitative with calibration standards
Mass spectrometry, since it can detect isoforms and discover
Olink, since it measures thousands of proteins at once
Western blot, since it uses a specific antibody per protein
Antibody and affinity methods measure predefined targets. Mass spectrometry measures peptides directly, can separate isoforms and supports discovery.
8.5 What a protein measurement means: units and NPX
| Kind | Meaning | Units |
|---|---|---|
| Absolute concentration | Real physical amount | mg/dL (US), mmol/L (EU), ng/mL or pg/mL for low-abundance proteins |
| Relative abundance | Relationships between samples are meaningful, the numbers are not physical concentrations | Often normalized, log-transformed, arbitrary units |
Platform units
- Olink and SomaScan: relative abundance as NPX (Normalized Protein eXpression) on a log₂ scale. One NPX unit = a doubling of signal (same logic as log₂FC in Lecture 7).
- Mass spectrometry: peptide intensity, ion counts or peak area; often calibrated by spiking in a labelled peptide of known concentration as an internal standard.
The first person has 3 times as much
The first person has 8 times as much
The first person has 3 mg/dL more
They differ by 3% in concentration
NPX is log₂: a difference of 3 NPX means 2³ = 8 times the relative abundance. It says nothing about absolute mg/dL.
8.6 QC, pre-analytical variation and batch effects
Pipeline: individual → blood sample → plasma/serum → assay → [QC, batch-effect handling, sample-prep issues such as freeze/thaw] → protein table (one row per person, one column per protein, e.g. Protein 1 … Protein 3000). Same shape as the GWAS table in Lecture 5, with continuous protein values instead of dosages.
QC at two levels
| Sample-level QC | Protein-level QC |
|---|---|
| Was the sample measured successfully? | Is the protein reliably detected across samples? |
| Is it an outlier across many proteins (a sample-wide problem)? | Too much missingness for this protein? |
| Too much missingness for this sample? | Is the assay reproducible? |
| Did internal technical controls fail? | Is its variation dominated by batch rather than biology? |
A very high or low value can be: real biology, contamination, technical failure, assay saturation (signal maxed out) or a processing error. Technical controls (for sample processing, hybridization/amplification, assay performance, normalization) help separate biology from artifact.
Pre-analytical variation (before the lab assay, often overlooked): fasting vs non-fasting, time of day, posture and recent exercise, tube/anticoagulant, time from draw to centrifugation, storage temperature, number of freeze–thaw cycles, hemolysis (red cells rupturing into the plasma), delay before freezing.
WarningSystematic, not random
Sick patients may be handled differently (more time, more controlled) than controls (processed faster, more randomly). Centers may have different protocols, and protocols change over time. If handling correlates with disease status, you get systematic bias.
Batch effects. A batch is any group of samples processed together that may share an artifact: the same plate, reagent lot, lab run, day, instrument or operator.
Detection: boxplots by plate/run (systematic shifts), PCA (do samples cluster by batch rather than biology?), clustering, technical controls, replicate samples across batches, comparing value distributions across plates.
| Prevent: design stage | Correct: analysis stage |
|---|---|
| Randomize samples across plates | Normalize across platforms/batches |
| Balance cases and controls across plates/batches | Adjust for plate/run as a covariate |
| Technical replicate samples across plates | Include batch as a covariate in the regression (like genotyping batch in GWAS, Lecture 4) |
| Common reference samples repeated across plates |
Transformations for proteomics data (right-skewed, wide dynamic range, sensitive to extremes):
- log transformation compresses the right tail;
- z-scores \(z = \dfrac{x - \text{mean}(x)}{\text{SD}(x)}\) (mean 0, SD 1) make effect sizes comparable across proteins on different scales;
- plate normalization;
- reference-sample normalization: the same sample on every plate, and all values on a plate normalized to it, removing plate-to-plate drift.
The study cannot use NPX units for relative abundance
Batch is completely confounded with case status
Too few proteins are measured on each individual plate
The protein values will no longer be right-skewed
Any plate effect is indistinguishable from the case–control difference. Randomizing and balancing cases and controls across plates at the design stage prevents this; no analysis step can fully fix it afterwards.
Part 2: PWAS and genetics–proteomics integration
8.7 The proteome-wide association study (PWAS)
Same structure as a GWAS, with a protein instead of a SNP as the predictor:
\[Y = \beta_0 + \beta_1 \times \text{protein} + \beta_2 \times \text{age} + \beta_3 \times \text{sex} + \dots\]
If the protein is standardized (z-score), \(\beta\) is the effect of a 1-SD higher protein level:
- continuous \(Y\): \(\beta = 0.18\) → 0.18 units higher outcome per SD;
- binary \(Y\) (logistic): \(e^{0.18} \approx 1.20\) → 20% higher odds per SD.
Two model directions
| Protein as predictor | Protein as outcome |
|---|---|
| \(Y = \beta_0 + \beta_1 \times \text{protein} + \text{covariates}\) | \(\text{protein} = \beta_0 + \beta_1 \times \text{exposure} + \text{covariates}\) |
| Does this protein predict the phenotype? | Does this exposure predict the protein level? |
Multiple testing in PWAS. ~3,000 proteins → about \(3{,}000 \times 0.05 = 150\) false positives at uncorrected \(p < 0.05\). In discovery, control the FDR (usually BH, Lecture 7). Unlike GWAS, there is no single conventional threshold like 5×10⁻⁸: it depends on the number of proteins, their correlation and the strategy.
8.8 Association is still not causation
Three explanations for a protein–disease association:
Causal diagrams (each arrow means “directly affects”): three different structures that all produce the same protein–disease association.
flowchart TD P1["protein"] --> D1["disease"]
flowchart TD D2["disease"] --> P2["protein"]
flowchart TD C3["confounder"] --> P3["protein"] C3 --> D3["disease"]
A protein can be a good biomarker/predictor without being causal. Troponin marks cardiac injury, but lowering troponin would not treat an MI: it is a readout of injury that already happened. Prediction and causation are different properties; only some proteins have both.
Troponin cannot be measured in plasma with current assays
It is released by injury, a consequence of the infarction
Troponin levels are fixed at conception like DNA
It is a trans-pQTL, so it has horizontal pleiotropy
Troponin is a readout of cardiomyocyte damage. Lowering it would not undo the damage: prediction without causation.
8.9 pQTLs: connecting genetics and proteomics
For a given SNP, do genotype groups (AA = 0, AG = 1, GG = 2) differ in protein abundance?
\[\text{Protein} = \beta_0 + \beta_1 \times \text{Genotype} + \varepsilon\]
(The SNP → continuous phenotype model of Lecture 4 with protein level as the outcome.) Questions it answers: which variants regulate which circulating proteins? Which genes influence protein levels? Can a protein be linked genetically to disease? Can the variant be an MR instrument?
pQTL (protein quantitative trait locus): a SNP associated with a protein’s abundance.
WarningA SNP can change function without changing abundance
Since assays measure amount, a functionally important variant can be invisible to a pQTL analysis if it does not also change the concentration.
PCSK9 chain: PCSK9 SNP → circulating PCSK9 protein → LDL receptor activity → LDL-C. DNA to clinical biomarker in four steps.
| cis-pQTL | trans-pQTL | |
|---|---|---|
| Location | Close to the gene encoding the protein | Far away: common working definition > 500 kb, or a different chromosome |
| Plausible mechanisms | Transcription, splicing, protein stability or secretion of that gene’s product | Indirect: transcription factors, receptors, signalling pathways, protein-processing machinery |
| Interpretation | Relatively direct biological interpretation | More indirect |
| As MR instrument | Preferred | Higher risk of horizontal pleiotropy |
A cis-pQTL, because it changes protein abundance
A trans-pQTL, because it is far from the encoding gene
A cis-pQTL, because any SNP near a gene counts as cis
Not a pQTL, because pQTLs must be on the same gene
cis = near the encoding gene; trans = far away (> 500 kb) or on another chromosome. Different chromosome → trans.
8.10 Protein Mendelian randomization
Genetic variant → protein abundance → disease risk. With \(\beta_{GP}\) = SNP effect on Protein and \(\beta_{GO}\) = SNP effect on Outcome:
\[\hat\beta_{MR} = \frac{\beta_{GO}}{\beta_{GP}}\]
The Wald ratio of Lecture 6 with protein level as the exposure: the effect on the outcome per genetically predicted unit change in protein.
Why cis-pQTLs are preferred as instruments: closer to the encoding gene, so an easier and more direct biological link, and less risk of horizontal pleiotropy (fewer alternative pathways than a variant acting at a distance). But cis-pQTLs still need all three MR assumptions (relevance, independence, exclusion restriction). Being cis reduces one risk; it does not remove the need for the other checks.
0.05 log-odds per SD of protein X
0.20 log-odds per SD of protein X
5.0 log-odds per SD of protein X
0.60 log-odds per SD of protein X
β_MR = β_GO / β_GP = 0.10 / 0.5 = 0.20 per SD, i.e. OR ≈ e^0.2 ≈ 1.22 per SD higher genetically predicted protein X.
8.11 Challenges specific to protein MR
| Challenge | Why it matters |
|---|---|
| Horizontal pleiotropy | Violates the exclusion restriction; especially relevant for trans-pQTLs |
| Assay binding artifacts | A coding variant can change the protein’s structure so the antibody/aptamer binds worse, giving a seemingly lower level with no real change in amount. Fairly unique to protein instruments |
| Amount ≠ function | A variant can change activity, binding, stability or localization without changing measured abundance |
| LD can fake causality | SNP A affects the protein, a different SNP B in LD affects disease through another mechanism: protein and disease look linked only through the LD |
| One pQTL, many proteins | SNP → protein A, B and C at once; the effect may not run through the one protein you study |
Colocalization: is it the same signal? Compare the pattern of association across a region for the protein’s pQTL and for the disease GWAS. If one shared causal variant drives both, SNPs in LD should give a matching pattern of association strength for both traits. If the patterns do not line up, the protein and disease associations are probably driven by two distinct (possibly nearby) causal variants. Check this before concluding that a pQTL and a GWAS hit at “the same locus” reflect the same biology.
The variant may change assay binding, not the amount
The variant must be a trans-pQTL since it is missense
The variant cannot affect protein function at all
The variant proves the protein causes the disease
Coding variants can alter the epitope the antibody or aptamer binds, producing an apparent drop in abundance: an assay binding artifact.
8.12 From association to confirmed mechanism
The lecture’s pipeline:
flowchart TD A["protein associated with disease"] --> B["pQTL evidence"] B --> C["MR and colocalization"] C --> D["candidate causal protein"] D --> E["mechanistic experiments"] E --> F["biological pathway"]
Follow-up questions once a protein is genetically implicated: Which cell or tissue produces it? Which receptor or pathway does it act through? Does changing its level or activity change cell function? Does it reproduce in an experimental model? Can the pathway be modified by a drug (is it druggable)?
Experimental validation approaches: CRISPR knock-out or knock-in; protein overexpression or targeted inhibition; cultured cells; organoids; animal models; receptor-binding experiments; enzyme activity assays.
The PCSK9 story, end to end (be able to reproduce it from memory)
| Step | |
|---|---|
| Genetics | PCSK9 genetic variation |
| Protein | PCSK9 protein abundance |
| Mechanism | PCSK9 promotes LDLR degradation → fewer LDL receptors on hepatocytes |
| Physiology | Reduced LDL clearance → higher LDL-C |
| Disease | Higher atherosclerotic risk |
| Drug (mirror image) | PCSK9 inhibition (Repatha/Praluent block the protein; Leqvio silences the mRNA) → more LDLR → lower LDL-C |
ImportantClosing statement of the course’s method arc
A protein can be statistically associated, genetically supported (pQTL/MR) and still need experimental validation before you understand the mechanism. This echoes Lecture 1: molecular epidemiology is not biology.
8.13 Reading material: eight plasma proteomics platforms on the same samples
(reading) Kirsher et al. (2025), “Current landscape of plasma proteomics from technical innovations to biological insights and biomarker discovery”, Communications Chemistry 8:279. This is the course’s reading material for Lecture 8. It puts numbers on the slides’ warnings: platforms measure different things, technical noise differs a lot, and “the same protein” on two platforms is often not the same measurement.
Design. Plasma from 78 healthy people (40 aged 55–65, 38 aged 18–22, half women) was measured on eight platforms: five affinity-based (SomaScan 7K and 11K with aptamers; Olink Explore 3072 and Explore HT with paired antibodies; NULISA, a small high-sensitivity antibody panel) and three mass spectrometry (MS) workflows (two discovery workflows that enrich low-abundance proteins with nanoparticles or remove high-abundance ones, and one targeted MS with internal standards, used as the reference because it gives absolute amounts).
Main results (numbers from the paper)
| What was compared | Finding |
|---|---|
| Coverage | 13,011 proteins in total; SomaScan 11K found the most (9,645). Only 36 proteins were measured by all eight platforms, 259 by all except NULISA |
| Technical CV (same sample measured repeatedly) | SomaScan about 5–6%, NULISA 6.6%, targeted MS 8.3%, Olink 3K 11.4%; Olink 5K 26.8% and the discovery MS workflows 26–30% |
| Data completeness (share of values above the detection limit) | SomaScan about 96%, Olink 3K 60%, Olink 5K 36% |
| Agreement between platforms (Spearman correlation for the same protein) | Two versions of one platform: about 0.75–0.8. Across technologies: often much lower, and bimodal: one group of proteins that agrees and one that does not |
| Age markers | Effect estimates of significant age proteins agreed well across platforms (correlation 0.66–0.95), much better than for all shared proteins (0.34–0.75) |
Four lessons that link to the rest of the course
- Technical vs biological variation (Lecture 7, 7.4). The paper compares the technical CV (replicates of the same sample) with the total CV across people. Where technical variation is a large share of the total, real biological differences are hard to see. Filtering Olink 5K to values above the detection limit lowered its CV from 26.8% to 12.4%, but removed about 40% of the analytes: precision is bought with coverage.
- Missing values and noise go together. Proteins with many values below the detection limit, usually low-abundance proteins, also had the highest CVs. Plasma spans about 10 orders of magnitude in concentration, which is why low-abundance proteins are hardest.
- Same name, different measurement. Affinity assays measure a binding epitope, not “the protein”. Most platforms measure total ApoE, but SomaScan’s “ApoE4” assay also binds ApoE2 and ApoE3 with similar affinity, so it is not isoform-specific; MS peptides unique to ApoE4 confirmed that NULISA’s ApoE4 assay is. This is the assay-binding issue of 8.11 in practice, and one reason why two platforms disagree.
- Multiple testing changes the hit list. Each protein was regressed on age, sex, BMI and other factors, with significance by raw p < 0.05 and by Benjamini–Hochberg adjusted p < 0.05 (Lecture 7, 7.13). How many markers survived BH depended on each platform’s p-value distribution, not only on how many proteins it measured.
Pre-analytical factors. The introduction lists what can shift plasma protein levels apart from the biology of interest: age, sex, BMI and fasting status, and technical factors such as storage time, temperature, time of day, anticoagulant and processing protocol. Nanoparticle enrichment added platelet contamination. This is 8.6’s QC and pre-analytical variation with concrete examples.
Take-home. No platform is best at everything: affinity platforms give the broadest coverage and best precision, targeted MS gives absolute quantification, discovery MS can see proteoforms. Choose by study goal, and treat a protein measured on one platform as that platform’s measurement, not as a universal value.
NoteRead critically: a detail in Table 1
The footnote of Table 1 says that the Wilcoxon rank-sum test was used for categorical variables and the chi-squared test for continuous variables. That is reversed: Wilcoxon is a test for continuous (or ordinal) variables, chi-squared for categorical ones. Probably a typo, but the kind of thing worth noticing when reading a paper.
The remaining proteins correlated worse with targeted MS
About 40% of the analytes dropped out of the comparison
Differences between people could no longer be estimated
Low-abundance proteins had to be remeasured on SomaScan
CV fell from 26.8% to 12.4%, but only because the noisy, mostly low-abundance analytes were removed: precision was bought with coverage.
Because one platform reports NPX and the other log₂ units
Because each binds its own epitope and may see other forms
Because plasma proteins change too fast to measure twice
Because correlation needs absolute, not relative, amounts
Affinity assays measure whatever their binder recognises. Different epitopes, isoform selectivity (the ApoE example) and binding artifacts make two assays for one protein into two different measurements.
Practice
NoteQuick calculations
Why-questions
TipWhy is a cis-pQTL a better MR instrument than a trans-pQTL?
A cis variant sits by the gene that encodes the protein, so its effect most plausibly runs directly through that protein (transcription, splicing, stability). A trans variant acts through other genes and pathways, so it more often affects the outcome by routes that bypass the protein: horizontal pleiotropy.
TipWhy is plasma level not the same as production?
Plasma level is a balance of synthesis, release, distribution, degradation, uptake and clearance. A kidney problem that slows clearance raises the level without any change in production.
TipWhy can a batch effect not simply be “corrected away” afterwards?
If batch is confounded with the comparison (all cases on some plates), batch and biology cannot be separated statistically. That is why randomization and balance at the design stage matter most.
TipWhy do we still need colocalization after finding a pQTL and a GWAS hit in the same region?
LD means two different causal variants close together can both produce signals at “the same locus”. Colocalization checks whether one shared variant explains both association patterns.
Spaced review
Inverse-variance weighting
Weighted median
MR-Egger regression
The Wald ratio
MR-Egger lets the intercept differ from zero; a non-zero intercept signals directional horizontal pleiotropy.
The probability of at least one false positive
The expected proportion of false hits among the hits
The variance of each feature across replicate arrays
The number of features that are truly differentially expressed
BH controls the FDR, the expected proportion of false discoveries. The probability of any false positive is the FWER (Bonferroni).
Exposure marker
Marker of tissue injury
Marker of future risk
Marker of inflammation
LDL-C is the course’s example of a marker of future risk; its change after statin therapy would be a treatment-response marker.
Further reading
Reading material
- Kirsher DY, Chand S, Phong A, Nguyen B, Szoke BG, Ahadi S (2025). Current landscape of plasma proteomics from technical innovations to biological insights and biomarker discovery. Communications Chemistry 8:279 (open access). Compares eight plasma proteomics platforms on the same 78 people. Summarised in 8.13.
In this app
- Martin’s key concepts: K.17 proteomics and protein abundance, K.18 pQTLs, K.19 colocalization, K.20 biomarker vs causal protein.
NoteCourse files on Canvas (GU login required)