# Codebook: a protein Mendelian randomization paper's headline claim

DRAFT, 2026-10-27. Readers code each paper alone. None sees another reader's codes, which papers
others coded, or any re-test result.

## The record

Three steps, each coded by two readers alone, with a third reader for their disagreements:

1. **Claim**, from the title and abstract as PubMed gives them: eligibility, the headline claim, and
   what the abstract says the paper found about it.
2. **Disease**, for each paper with a claim, from its title and abstract and its settled claim: the
   disease's name, its Open Targets term, and its ICD-10 code.
3. **Data sources**, for each paper with a claim, from its abstract and, where Europe PMC holds it, its
   full text: where the protein's instruments and the disease's summary statistics came from, their
   scope and ancestry, and the unit of the paper's estimates.

## Step 1: eligibility and claim

1. **eligibility:** `eligible`, or the first criterion of the plan's section 4 the paper fails,
   checking E0, E1, E2, E3, E4, and E5 in that order, with a sentence quoted from the record that
   shows it.
   - A paper meets E2 if any of its MR analyses uses variants associated with a circulating protein's
     level as that protein's instruments, even if its other analyses use other exposures. A protein
     measured as a clinical biomarker (apolipoprotein B, C-reactive protein) is a circulating protein
     like any other; what's biomarker-weighted is a drug target's variants weighted by such a biomarker.
     Lipoprotein(a) is apolipoprotein(a), the LPA gene's product. When it fails
     E2, **e2_kind** says how: `expression or other molecular QTLs` (gene expression, methylation,
     splicing), `not circulating` (cerebrospinal fluid, brain, or other tissue proteins),
     `biomarker-weighted` (a drug target's variants weighted by a downstream biomarker, such as PCSK9's
     by LDL cholesterol or IL6R's by CRP), or `no protein exposure` (the exposure isn't a protein, or the
     MR runs from the disease to the protein), or `not one gene's product` (a complex, a ratio, a
     score).
   - An exposure the abstract names as the protein's circulating level ("genetically predicted plasma
     C2", "per SD of kit ligand", a proteome-wide MR of plasma proteins) is instrumented by variants
     associated with that level, so it meets E2 whether or not the abstract says "pQTL". `E2 unstated`
     is for an exposure that could have been instrumented otherwise and whose instruments the abstract
     doesn't name: a protein's "signalling", "inhibition", or "activity", or "variants in the gene";
     and for a pair from analyses that pool kinds of QTL (eQTLs and pQTLs in one SMR) or tissues
     (circulating with cerebrospinal fluid or brain pQTLs), or that rest on "expression" not named as a
     gene's, an mRNA's, or a protein's, unless the abstract ties the pair to the circulating protein's
     analysis. Judge the instruments only from what the abstract says (a target "proxied by" or
     weighted by a named biomarker is biomarker-weighted); don't infer them from hints. Code its claim as
     for an eligible paper; step 3 settles E2 from the methods.
   - A methods paper whose analysis estimates no protein's effect on a disease fails E2, not E5; a paper
     whose instruments are pQTLs but whose outcome is a trait fails E3, not E4.
   - E3: an outcome analysed as a continuous score, a frequency, or a severity (a self-reported symptom's
     frequency), a serostatus, or a measurement isn't a disease.
   - E4: the abstract states no protein-disease pair from the protein's MR (below), with or without a
     direction.
2. **claim**, for an eligible (or `E2 unstated`) paper: the headline pair.
   - **Which pairs count.** A protein-disease pair the abstract attributes to MR of the protein's
     circulating level on the disease, or any MR pair in a paper whose only MR exposures are circulating
     proteins. Pairs from gene expression, methylation, CSF or brain proteins, variants weighted by a
     biomarker, reverse MR (disease to protein), or observational or experimental data don't count, even
     when restated as the protein's effect. Nor does a pair the abstract says didn't survive correction
     for multiple testing. A protein the abstract names as passing a filter applied to its MR results
     (colocalisation, replication, sensitivity analyses) is a counting pair with the disease. An
     association of protein-altering (missense) variants with the disease isn't MR of the protein's
     level. A protein named as mediating another exposure's effect is a counting pair with the disease
     only where the abstract gives the protein's own effect on it (the mediator's step of the MR). A
     class of diseases ("autoimmune diseases") stands for the first member disease the abstract pairs
     with the protein; with none, the pair doesn't count. A sentence that lists proteins and diseases pairs them only where it says
     each protein holds for each disease; the pair is then its first protein with its first disease.
   - **Which pair.** The first counting pair the conclusions state with a direction; if the
     conclusions name proteins but give none a direction, the first of those whose direction the
     results state; if they name none, the first counting pair the results state with a direction; and
     if no counting pair has a direction, the first counting pair the conclusions name, else the first
     the results name, with the direction `unstated`.
   - **What a direction is.** Words that say which way higher protein moves risk ("a risk factor",
     "protective", "increased the risk", "inhibiting X may prevent Y"). Words that say only that a
     protein is causal, involved, contributes, mediates, or is pathogenic give no direction. A protein
     named only as mediating another exposure's effect has a direction only where the abstract states the
     sign of its own effect on the disease. A direction stated for a group of proteins ("16 proteins
     showed inverse causal relationships") is each named member's. A direction from observational or
     experimental data, even beside the MR, isn't the MR's. An odds ratio overrides words.
   - **protein_as_written:** the protein exactly as the abstract writes it.
   - **gene:** the HGNC symbol of the gene encoding it (PCSK9 → PCSK9; IL-6 receptor → IL6R; MSP →
     MST1; TL1A → TNFSF15; sRAGE → AGER). A protein that isn't one gene's product (a complex, a
     ratio, a score) fails E2, of the kind `not one gene's product`.
   - **disease_as_written:** exactly as the abstract writes it.
   - **direction:** `raises` if higher genetically predicted protein raises the risk, `lowers` if it
     lowers it, `unstated` if no counting pair has a direction.
   - **estimate:** the odds ratio and its 95% interval as given, with the unit as the abstract states
     it (per standard deviation, per unit, other, or not stated), and its **frame**: `per higher
     protein`, or `per lower protein` where it is given per lowering or inhibiting the protein (an OR of
     0.84 for "inhibiting X" with the direction "raises"); null if the abstract gives none.
   - **significance:** `corrected` (the abstract says the pair survived correction for multiple
     testing, by FDR, Bonferroni, or another, or gives a threshold that is a significance level divided
     by the number of tests, such as p < 0.05/110), `nominal` (it gives only an uncorrected p-value, or
     calls the pair nominally significant or suggestive), or `not stated`.
   - **quote:** the sentence that states the claim, verbatim.
3. **For the claim**, each with its quote when true:
   - **reports_colocalisation:** the abstract says the headline protein colocalised with the disease,
     or gives its colocalisation probability; **coloc_criterion:** the criterion it states (`PP.H4`,
     `PP.H3 + PP.H4`, `other`, or `not stated`); **coloc_value:** the headline protein's probability as
     given (0.721 for "PP.H4 = 72.1%"), else null.
   - **reports_replication:** the abstract names a second outcome cohort in which the headline pair, for
     the same disease, was analysed separately. A replication of the pQTLs, a test in a meta-analysis
     that holds the discovery cohort, or "validated" without a named cohort isn't one.
   - **drug_target_framing:** the abstract names the headline protein as a drug or therapeutic target,
     or names a drug acting on it. Calling the paper's findings in general "potential targets" isn't
     enough.

## Step 2: disease

For each paper with a claim, given its settled claim:

- **disease_name:** the claim's disease by its full name, as the abstract defines it: an abbreviation
  expanded as the abstract expands it (in a paper on pancreatic cancer, "PC" is pancreatic cancer). If
  the claim names more than one disease ("kidney and bladder cancer"), the first one named.
- **disease_id:** the Open Targets 26.09 disease (EFO, MONDO, HP, or Orphanet ID) whose name or exact
  synonym is the disease; if several, the narrowest that still covers the paper's disease (for
  "depression", depressive disorder, not major depressive disorder); if none matches, the nearest
  broader one, with **broader** true. Look the disease up by its full name, and
  also by its other and former names (NAFLD is now MASLD; premature ovarian failure is primary ovarian
  insufficiency): a broader term is the nearest one that still names the disease's own kind, never a
  whole organ's or system's diseases when a term for its kind exists. Prefer a disease term (MONDO,
  EFO) to a phenotype term (HP) that names the disease's recurrence or a sign of it ("recurrent upper
  respiratory tract infections" for upper respiratory tract infection). Open Targets' names can
  mislead: take the term whose synonyms name the disease (multiple sclerosis is EFO_0803536).
- **icd10:** the ICD-10 codes whose union is the disease: three-character categories (type 2 diabetes:
  E11; inflammatory bowel disease: K50 and K51), and four-character subcategories where the disease is
  part of a category (periodontitis: K05.2 and K05.3; attention deficit hyperactivity disorder: F90.0;
  abdominal aortic aneurysm: I71.3 and I71.4; Hirschsprung disease: Q43.1); with **icd10_broader**
  true when the codes cover more than the disease (small vessel stroke: I63, all cerebral infarction),
  or an empty list when no code fits.

## Step 3: data sources

For each paper with a claim, given its settled claim, from its abstract and its full text where Europe
PMC holds it (the methods, and the results that say which analysis gave the claim), each with the
sentence that says it, verbatim:

- **instruments:** whether the claim's protein's instruments are variants associated with its
  circulating level, in plasma or serum (`pQTLs`), something else (`other`, with **instruments_kind**
  one of step 1's e2_kind values, cerebrospinal fluid or brain pQTLs being `not circulating`), or `not
  stated`.
- **pqtl_source:** every source of the claim's protein's instruments, each one of deCODE, UKB-PPP,
  Fenland, ARIC, INTERVAL, SCALLOP, other (named in **pqtl_other**), or not stated; **platform:** each
  platform they were measured on, Olink, SomaScan, other, or not stated, as the paper names it. A source
  the paper describes without naming is the standard one its description fits:

  | Source | Its description |
  |---|---|
  | deCODE (Ferkingstad 2021) | 35,559 Icelanders; 4,907 aptamers or 4,719 proteins; SomaScan v4 |
  | UKB-PPP (Sun 2023) | 54,219 UK Biobank participants; 2,923 proteins; Olink Explore 3072 |
  | Fenland (Pietzner 2021) | 10,708 participants; 4,775 protein targets; SomaScan |
  | ARIC (Zhang 2022) | 7,213 European-ancestry participants; 4,657 proteins; SomaScan |
  | INTERVAL (Sun 2018) | 3,301 participants; 3,622 proteins; SomaScan |
  | SCALLOP | Olink panels' consortium: 30,931 participants and 90 cardiovascular proteins (Folkersen 2020), or 14,824 and 91 inflammatory proteins (Zhao 2023) |

  A collection of several studies' pQTLs (Zheng 2020's 734 plasma and 154 cerebrospinal fluid proteins)
  is `other`, named in pqtl_other.
- **instrument_scope:** the claim's instruments as the paper chose them: `cis` (variants near the
  protein's gene), `trans`, `cis and trans`, `genome-wide` (every significant variant), or `not stated`.
- **ancestry:** the ancestry of the outcome data that gave the claim's estimate: `European`, `East
  Asian`, `South Asian`, `African`, `other`, `multi-ancestry`, or `not stated`.
- **or_unit:** the unit of the paper's MR estimates as the abstract or full text state it: `per
  standard deviation` of the protein, `per unit`, `other`, or `not stated`.
- **outcome_sources:** every source of the claim's disease's summary statistics the paper used for the
  claim's MR, in discovery or replication; not the data of the exposure, a mediator, or another
  analysis: each with its **name** (FinnGen, UK Biobank, a named consortium or study),
  **release**, and **accession** (a GWAS Catalog GCST accession or IEU OpenGWAS ID, as given), and
  whether it **includes_finngen** and whether it **includes_ukb**: `yes`, `no`, or `unknown`, with the
  **basis** for each that isn't unknown:
  - `stated`: the paper names the cohorts the source holds;
  - `identifier`: the source's ID names it (an OpenGWAS `finn-b-` ID is FinnGen, a `ukb-` ID is UK
    Biobank), or the source is FinnGen or UK Biobank itself;
  - `published composition`: the named consortium's or study's own publication lists its cohorts
    (HERMES's 2020 meta-analysis of heart failure includes UK Biobank; IGAP's 2019 meta-analysis of
    Alzheimer disease doesn't), as the reader knows it.

  Code `unknown` when none of these settles it. A source the paper names without saying what it holds,
  and whose make-up the reader doesn't know, is unknown for both. A source holding Finnish cohorts
  (FINRISK, Health 2000, METSIM, the Botnia study, and others) that FinnGen may also hold is `unknown`
  for FinnGen unless the paper settles it.

## Output

One JSON object per paper and step, in the order given (the values are an example, not a paper):

```json
{"pmid": "00000000", "eligibility": "eligible", "e2_kind": null, "eligibility_reason": "...", "eligibility_quote": "...",
 "claim": {"protein_as_written": "PCSK9", "gene": "PCSK9", "disease_as_written": "CAD", "direction": "raises",
           "estimate": {"or": 1.21, "low": 1.10, "high": 1.33, "unit": "not stated", "frame": "per higher protein"}, "significance": "corrected",
           "quote": "..."},
 "flags": {"reports_colocalisation": true, "coloc_criterion": "PP.H4", "coloc_value": 0.92, "colocalisation_quote": "...", "reports_replication": false,
           "drug_target_framing": true, "drug_target_quote": "..."}}
{"pmid": "00000000", "disease_name": "coronary artery disease", "disease_id": "MONDO_0005010", "broader": false,
 "icd10": ["I20", "I21", "I22", "I23", "I24", "I25"], "icd10_broader": false}
{"pmid": "00000000", "instruments": "pQTLs", "instruments_kind": null, "instruments_quote": "...", "instrument_scope": "cis",
 "ancestry": "European", "pqtl_source": ["UKB-PPP"], "pqtl_other": null,
 "platform": ["Olink"], "pqtl_quote": "...", "or_unit": "per standard deviation", "or_unit_quote": "...",
 "outcome_sources": [{"name": "FinnGen", "release": "R10", "accession": null, "includes_finngen": "yes",
                      "finngen_basis": "identifier", "includes_ukb": "no", "ukb_basis": "identifier", "quote": "..."}]}
```

Quotes are for checking codes against the record and are never published; the bundle keeps only their
character offsets.

## Settling disagreements

The two readers' codes are compared on: in step 1, eligibility, gene, direction, the disease as
written (ignoring case, punctuation, and white space), the odds ratio and its interval, significance,
and the three flags; in step 2, disease_id, icd10, and icd10_broader; in step 3, instruments,
pqtl_source, platform, or_unit, instrument_scope, ancestry, and, from the sources, whether the original's outcome data include
FinnGen and whether they include UK Biobank (yes if any source does, no if every source doesn't,
unknown otherwise). Each disagreement goes, with the record and both codes but without saying which
reader wrote which, to a third reader, who decides it under this codebook. Quotes are checked against
the record by string matching, after Unicode normalization (NFKC) and with runs of white space read as
one space; a quote that doesn't match is a disagreement, decided by the third reader.
