# ABM PGx Atlas — catalogue of computed parameters Per-variant and per-association parameters available in the resource. Unit of analysis: **variant × drug × ICD-10 condition** ("association"). --- ## 1. Study design Each association is characterised against **four independent GWAS phenotype arms** run on the same biobank cohort, which is what makes the contrast parameters (§3) possible: | arm | phenotype | model | |---|---|---| | **ADR (discovery)** | condition occurring in users of the drug | logistic (REGENIE) | | **disease-alone** | the ICD-10 condition in the general population, drug-free | logistic | | **drug dose** | `__median_dose__ln` | **linear** | | **drug prescribed** | propensity to be prescribed the drug | logistic | Cohort is split **80% discovery / 20% held-out test**; the discovery arm is run on the 80%, and every association carries an independent replication statistic from the held-out 20% (§4). --- ## 2. Association statistics — per arm For each of the four arms: - `BETA` — effect size (log-odds; linear for the dose arm) - `SE` — standard error - `LOG10P` — −log10 p-value - `CHISQ` / `Z_STAT` — test statistics --- ## 3. Contrast parameters (differences of betas) The most informative family — they isolate what is **specific to the drug** rather than to the condition or to the patient population: - `delta_BETA_ADR_vs_pure` — ADR effect vs the same variant's effect on the disease alone - `delta_BETA_ADR_vs_dose` — ADR effect vs the dose-response arm - `delta_BETA_ADR_vs_prescribed` — ADR effect vs prescription propensity - `delta_BETA_pure_vs_prescribed` — disease-alone vs prescription propensity - `beta_superiority` — ADR effect relative to the **strongest** drug-free arm, SE-shrunk and frequency-standardised; untested arms excluded rather than treated as null - `effect_sign_concordance` — sign agreement between arms - `drug_specificity` — how much of the signal is drug-conditional Betas are normalised across arms before contrasting (the dose arm is linear, the others logistic), and standardised per-allele by `β·√(2f(1−f))` to remove frequency-driven inflation in rare variants. --- ## 4. Held-out replication - `repl_directional_z` — `sign(BETA_discovery) · log(OR_test) / SE_test`; one-sided, precision-weighted, direction-aware - `OR_logdiff_test_minus_disc`, `OR_ratio_test_over_disc` — test-vs-discovery effect calibration - `test_Z_STAT`, `LOG(OR)_SE` — held-out test statistics - confidence tier — replicated / replication-underpowered / unreplicated Directional replication is used rather than a p-value threshold because most candidates are rare; 201/207 are sign-concordant, where only 105 would pass a two-sided P<0.05. --- ## 5. Genotype and frequency - `A1_CASE_CT`, `A1_CTRL_CT`, case/control allele counts - `A1_CASE_FREQ`, `A1_CTRL_FREQ` — allele frequency in cases and controls separately - `CASE_HET_A1_CT`, `CASE_HOM_A1_CT`, `CTRL_HET_A1_CT`, `CTRL_HOM_A1_CT` — genotype counts - `A1_FREQ`, `INFO` (imputation quality), `N` - `gnomAD_AF` + missingness indicator --- ## 6. Temporal / EHR-derived (ADR-specific) Derived from prescription and diagnosis dates — these describe the **time relationship between drug exposure and the event**, and are among the most discriminating parameters we have: - `TEMP_n_before_exposure`, `TEMP_n_after_exposure` — event counts either side of first exposure - `TEMP_pct_before`, `TEMP_pct_after` — proportion of diagnoses before / after the prescription - `ADR_after_over_before` — directional log-ratio (drug-induced vs pre-existing direction) - `TEMP_median_days_after`, `TEMP_median_years_after` - `TEMP_median_days_before_exposure`, `TEMP_median_years_before_exposure` - `TEMP_q25_days_after`, `TEMP_q75_days_after` — onset-window location (tight window = drug-locked) - `TEMP_pct_missing_dx_date`, `TEMP_pct_missing_rx_date` — data-quality indicators Empirically: events following exposure at ~30× the pre-exposure rate is the drug-effect signature; a high pre-exposure fraction (~15%) is the indication signature. --- ## 7. Pharmacovigilance (FAERS) Real-world spontaneous-report disproportionality for the same drug–event pair: - `FAERS_PRR` (+ CI), `FAERS_ROR` (+ CI), `FAERS_IC025`, `FAERS_a` (report count) - `FAERS_MedDRA_PT` — mapped preferred term Caveat we document: these are elevated for **on-label** pairs (confounding by indication), so they corroborate reporting, not causality. --- ## 8. Variant annotation - VEP `Consequence` (+ canonical severity rank), `IMPACT`, `BIOTYPE`, `SYMBOL`, `Gene`, `Feature` - `SIFT`, `PolyPhen` scores - `Existing_variation` (rsID), position, REF/ALT - `internal_pleiotropy` — number of hits for the variant across our own GWAS corpus (computed internally, de-contaminated of ancillary biobank traits) --- ## 9. Label / indication context - `BNF_indication_codes`, `BNF_indication_description` — licensed indications for the drug - `on_BNF` — whether the ICD-10 condition is on the drug's own label --- ## 10. Composite / derived outputs - **latent-PGx priority score** — equal-weight geometric mean (soft-AND) of five evidence axes: directional replication, FAERS disproportionality, temporal direction, effect superiority, and inverse pleiotropy. No tuned weights. - **five-state mechanistic classification**, assigned from parameter signatures alone (unsupervised — no expert labels): | state | meaning | |---|---| | **I** | Indication — drug given *for* the condition | | **E** | drug-triggered Effect — drug-specific, follows exposure, replicates | | **P** | Preexisting effect enhanced by the drug | | **C2** | Confounder — a locus drives the disease/prescription, the drug adds nothing | | **C** | Comorbidity — weak, nonspecific co-occurrence | - `evidence_stouffer_Z`, `evidence_conjunction_minZ` — cross-arm evidence combination (Stouffer 1949; Nichols 2005 conjunction) - cluster assignment + per-cluster mechanistic label --- ## 11. Scale - **44** parameters in the classification model, derived from ~**164** source columns - **207** locus × phenotype associations after collapsing co-located variants to lead signals (250 kb window per drug × condition) - 4 GWAS arms × ~2.7 M variants in the underlying summary-statistics corpus --- ## What is distinctive here 1. **Four-arm contrast design** — the same variant is measured with the drug, without the drug, against dose, and against prescription propensity. This is what separates a drug-triggered effect from a predisposition or a confounder. 2. **Temporal EHR parameters** — event timing relative to first exposure, which pure summary-statistics resources do not carry. 3. **Held-out directional replication** on an independent 20% of the cohort. 4. **Pharmacovigilance linkage** to the same drug–event pair.