# Intellectual assessment: are the curated targets true drug effects or comorbidity? Literature-based review of the 6 curated target associations and the 15 highest-CSI associations. Each is judged on three axes: (1) does the drug **pharmacologically** cause this event, (2) is the **direction** right, (3) is the **gene** plausible for the mechanism. --- ## Part 1 — the 6 curated targets ### 1. gabapentin x R42 "Dizziness and giddiness" — ELOVL5 — **TRUE TARGET (best in set)** Dizziness/ataxia is the hallmark labelled adverse effect of gabapentin. `ELOVL5` mutations cause **spinocerebellar ataxia 38** — cerebellar Purkinje-cell degeneration with adult-onset ataxia (Am J Hum Genet 2014; Hum Genet 2023). A variant in a cerebellar-ataxia gene modulating susceptibility to a drug whose principal toxicity is ataxia is a genuinely coherent pharmacogenetic hypothesis — drug mechanism and gene mechanism converge on the same tissue. *Caveat:* our variant-level support is weak (1 case carrier, Replicability −1.4). Strong biology, thin statistics. ### 2. citalopram x M81.9 "Osteoporosis" — DLEU1 — **TRUE ADR (class effect), gene weak** Well-established: SSRI users have **1.61x** the fracture rate and **1.6x** the rate of bone loss; serotonin transporters are present on osteoblasts, osteoclasts and osteocytes, giving a direct mechanism (rather than a purely behavioural one). *Caveats:* depression itself lowers bone mineral density, so part of the association is confounding by indication; 13.6% of diagnoses precede the prescription. `DLEU1` is a lncRNA in the 13q14 CLL-deletion region with no bone biology — the gene link is unsupported. ### 3. ramipril x R41.0 "Disorientation" — GJB6 — **TRUE ADR pharmacologically, variant NOT supported** ACE inhibitors cause SIADH-type dilutional hyponatremia; sodium as low as 101 mEq/L produces confusion through to seizures and coma, and "confusional state / disorientation" appear as documented serious adverse events for ramipril. *But our data does not support this variant:* **0 test-set carriers**, OR 0.99, Replicability **−2.4**. `GJB6` is connexin-30 (a deafness gene) with no plausible route to hyponatremic confusion. The drug effect is real; **this association is an artifact**. ### 4. diazepam x J44.9 "COPD" — CXCL13 / PDE1C — **LIKELY COMORBID / reverse causation** Benzodiazepines genuinely harm COPD patients: **+45% exacerbations**, more respiratory failure and increased short-term mortality. But benzodiazepines do **not cause COPD** — COPD is a smoking disease. An incident J44.9 code after diazepam is therefore most plausibly either an exacerbation coded as COPD, or confounding by indication (breathless, anxious COPD patients are prescribed benzodiazepines). *Interesting:* `CXCL13` drives lymphoid-follicle neogenesis in COPD lungs, so the gene is not implausible. But the causal arrow for an *incident diagnosis* most likely runs the wrong way. **Treat as comorbidity, not a drug-triggered ADR.** ### 5. lisinopril x R73.9 "Hyperglycaemia" — RYR2 — **CONFOUNDER (reverse pharmacology)** This is the clearest negative. ACE inhibitors **improve** insulin sensitivity, **lower** glucose, and **reduce** new-onset type 2 diabetes; they can even cause **hypo**glycaemia. Hyperglycaemia as a lisinopril adverse effect is therefore **pharmacologically backwards** — the identical signature to `simvastatin x I83.1` (statins protect veins). The mechanism is confounding by indication: lisinopril is prescribed for diabetic nephropathy and for hypertension in diabetics, so lisinopril users are enriched for diabetes. 11.5% of hyperglycaemia diagnoses precede the prescription. *Note:* `RYR2` does have a beta-cell insulin-secretion role, so a variant-specific effect is not impossible — but the population-level direction contradicts the drug's pharmacology. **Recommend demoting to confounder, alongside simvastatin x I83.1.** ### Summary of the six | association | verdict | why | |---|---|---| | gabapentin x R42 | **true target** | ataxia drug x ataxia gene (ELOVL5/SCA38) | | citalopram x M81.9 | **true ADR**, weak gene | SSRI bone loss is established; DLEU1 implausible | | ramipril x R41.0 | real ADR, **artifact variant** | ACE-i hyponatremia real; 0 carriers, repl −2.4 | | diazepam x J44.9 (x2) | **likely comorbid** | benzos worsen COPD but do not cause it | | lisinopril x R73.9 | **confounder** | ACE-i *lowers* glucose — direction reversed | Only **one** of the five pairs is a clean drug-triggered target on both pharmacology and genetics. --- ## Part 2 — the 15 highest-CSI associations ### Pharmacologically credible (worth following up) | rank | association | gene | assessment | |---|---|---|---| | 1 | simvastatin x G62.9 polyneuropathy | PLB1 | **Contested but real.** Case-control studies report 4-14x risk; a 2021 meta-analysis finds none; attributable risk ~12/100,000 person-years. A legitimate open pharmacovigilance question — a good #1. | | 7 | fluoxetine x R55 syncope | CDH22 | **Strong.** SSRI-induced SIADH hyponatremia causes unsteadiness, falls and syncope, concentrated in the first 1-3 weeks and worse in the elderly. Mechanistically clean. | | 12 | amlodipine x R06.0 dyspnoea | GRM8 | **Plausible.** Amlodipine causes dose-dependent peripheral edema (10.8% at 10 mg) and dyspnoea appears in FAERS analyses of amlodipine. | | 11 | atorvastatin x K30 dyspepsia | SEL1L3 | Plausible but mild — statin GI intolerance is common and low-stakes. | | 6 | amlodipine x E86 volume depletion | XPO6 | **Direction questionable** — amlodipine causes fluid *retention*, not depletion. More likely co-prescribed diuretics. | ### Non-specific codes with no obvious pharmacology (likely noise) Ranks 2-5, 8-10, 13-15: citalopram x gastric polyp, ramipril x polyuria, ramipril x pleural effusion, citalopram/ramipril x gastroenteritis (A09.9, an *infectious* code), bisoprolol x acidosis, simvastatin x dysphagia, ramipril x gout, simvastatin x trochanteric bursitis, ramipril x tinnitus. These are mostly non-specific R-codes and A-codes with no documented drug mechanism. They are the expected background of a rare-variant scan. ### One methodological check, run and **refuted** I suspected CSI's top was inflated by tiny event counts (a 0% pre-exposure fraction is trivial when there are 2 events). **It is not.** The top-20 rests on *more* evidence than the atlas median: 176 vs 126 temporal events, 2.0 vs 1.0 case carriers, none with <=5 events. CSI correlates only weakly with event count (rho +0.08) and carrier count (rho +0.21), so it is not a disguised power score. The suspicion is recorded here because it was tested and rejected, not assumed. --- ## What this means 1. **The curated target list does not survive scrutiny.** Of the five pairs, one is a clean target (gabapentin x R42), one is a real class effect with an implausible gene (citalopram x M81.9), one is a real ADR on a statistically unsupported variant (ramipril x R41.0), one is likely comorbidity (diazepam x J44.9), and one is a reverse-pharmacology **confounder** (lisinopril x R73.9). This independently corroborates §9.15: CSI ranked them mid-table because the evidence really is mid-table. 2. **A second confounder has been identified.** `lisinopril x R73.9` has the same signature as `simvastatin x I83.1` — an event in the direction *opposite* to the drug's established pharmacology, in a population enriched for the event by the drug's indication. It is a candidate for `KNOWN_CONFOUNDERS`. 3. **CSI's own top hits are more defensible than the curated list.** Simvastatin x polyneuropathy and fluoxetine x syncope are both mechanistically coherent and match documented pharmacology, and neither was on anyone's list. 4. **The generalisable pattern:** the confounders in this atlas are identifiable by *reverse pharmacology* — the drug is known to protect against, or is prescribed because of, the event. That is a pair-level literature fact, not a variant statistic, which is exactly why no parameter in the model can capture it (§9.8-9.11).