ABM Atlas — Literature Validation

Expert cascade (EPIC) vs unsupervised classifiers (k-means / GMM / LCA / HDBSCAN)

v20 model · June 2026

Purpose

Ten randomly-selected associations plus two Effect (E) examples were checked against the published scientific literature, to judge whether the expert cascade classification and the unsupervised cluster assignment are biologically accurate. EPIC: E = Effect (drug causes the ADR), P = Predisposition, I = Indication (confounding by indication), C = inConclusive (low-evidence residual).


Part A — 10 random associations

1. simvastatin × varicose veins (I83.1) × DUSP7 — Expert E vs unsupervised P / P / P / I

The expert cascade calls this Effect (drug causes the event) on FAERS=STRONG, tissue=enriched, N_INDEPENDENT=3. The literature points the opposite way: statins suppress varicose remodeling — HMG-CoA reductase inhibitors block varicose vein development in mice, atorvastatin/rosuvastatin inhibit wall-stress AP-1 activity in venous smooth-muscle cells, and simvastatin has been trialled for venous ulcers. The curator agrees: “not an indication; shared vascular risk factors (age, obesity, sedentary lifestyle); weak comorbidity.” DUSP7 (a dual-specificity phosphatase) has no venous mechanism.

This is the clearest case where the expert classification is likely wrong: labelling a protective/confounded association as a drug-caused Effect inverts the biology. The unsupervised methods, pulling it away from Effect toward Predisposition (3/4) or Indication (HDBSCAN), show appropriate skepticism — though none lands on the most defensible label, inConclusive/comorbidity (shared metabolic risk driving both statin use and varicose veins). Verdict: expert E is biologically backwards; true class is confounding (C).

2. atenolol × psoriasis (L40.9) × MICAL3 — Expert C vs unsupervised P / P / P / P

Beta-blockers including atenolol are historically among the drugs most strongly tied to drug-induced/aggravated psoriasis (~20% of psoriasis patients; calcium/cAMP keratinocyte mechanism), with atenolol prominent in adverse-event reports (Cohen BJD 2008; Brauchli BJD 2008, OR 1.31). However, a 2022+ multi-method reanalysis (NHANES + FDA) found no significant signal, arguing the label may be inappropriate. So this is a real-but-contested drug-induced effect.

The expert calls it Comorbidity — but if atenolol genuinely induces psoriasis, that is an Effect, not coincidental co-occurrence. The cascade is internally inconsistent: the curator quotes strong causal evidence yet the tier lands on the weakest causal class. All four unsupervised methods say Predisposition — closer to “drug-related” than the expert’s Comorbidity. Verdict: expert undercalls a plausible drug-induced ADR; unsupervised P is more defensible, though the honest answer is “contested.” MICAL3 is almost certainly incidental.

3. lisinopril × oesophagitis (K20) × GNG13 — Expert C vs unsupervised P / P / P / P

ACE-inhibitor cough is a well-known class effect that can clinically mimic GERD/oesophagitis — the “diagnostic confounding” the curator flags. There is no mechanism by which lisinopril causes true oesophageal inflammation, and GNG13 (a G-protein γ-subunit in taste/neural tissue) has no oesophageal role; the signal is weak (FAERS weak, tissue low, N_INDEPENDENT=1).

Here the expert Comorbidity/inConclusive call is the more defensible one — it correctly reads the association as confounding rather than a real drug mechanism. The unanimous unsupervised “Predisposition” reflects the systematic bias: the rare-variant axis (GNG13 is a rare variant) pulls these signals toward P regardless of biology. Verdict: expert right, unsupervised over-calling.

4. amlodipine × pain, unspecified (R52.9) × NPAS3 — Expert P vs unsupervised P / P / P / I

The outcome is the problem: “unspecified pain” is maximally non-specific, and NPAS3 is a neurodevelopmental transcription factor (brain-expressed) with no link to amlodipine or pain. The Predisposition call was driven by tissue=“specific” (a brain-expression artifact), with otherwise weak evidence.

Neither classification is trustworthy because the association is near-noise — a vague outcome plus an incidental rare variant. Expert P and three unsupervised P agree, but on thin grounds; HDBSCAN’s dissent to Indication is no better supported. Verdict: low confidence on both sides; should be flagged uninterpretable.

5. atorvastatin × hypertrophic skin disorders (L91.8) × GALNT12 — Expert C vs unsupervised P / P / P / P

L91.8 (keloids and other hypertrophic skin disorders) has no established statin association, and GALNT12 (a GalNAc-transferase, colorectal-cancer gene) has no skin-hypertrophy mechanism. Evidence is weak throughout; the curator notes “no indication.”

The expert Comorbidity/inConclusive is appropriate — it reads a weak, mechanism-less association as low-confidence co-occurrence. The unanimous unsupervised Predisposition is again the rare-variant pull misfiring. Verdict: expert correct, unsupervised systematically over-calling P.

6. ramipril × disorientation (R41.0) × GJB6 — Expert P vs unsupervised P / P / P / I

This has a genuine pathway: ACE inhibitors can cause hyponatremia/SIADH, and severe hyponatremia (Na⁺ <120 mEq/L) produces confusion/disorientation in the elderly — documented in case reports where confusion resolved on drug withdrawal. So ramipril can cause disorientation, but indirectly via electrolytes, not through GJB6 (connexin-30, a deafness/skin gene). Signal moderately strong (FAERS=STRONG, tissue enriched, N_INDEPENDENT=2).

The drug→ADR link is real, making the expert Predisposition plausible (arguably edging to Effect). But the gene-level claim is spurious — the causal chain is drug→hyponatremia→confusion, independent of GJB6. Verdict: drug-ADR plausible (P defensible), gene incidental; strong FAERS justifies taking the signal seriously while distrusting the variant.

7. atorvastatin × fatty liver (K76.0) × FBLIM1 — Expert I vs unsupervised P / P / P / I

Statins are not formally indicated for NAFLD but are safe and beneficial (atorvastatin reduces aminotransferases, ameliorates NASH; recommended for the accompanying dyslipidemia). NAFLD is a strong metabolic-syndrome marker, so statin users are enriched for fatty liver: confounding by (metabolic) indication.

The expert Indication call is the right read — it identifies confounding-by-indication even if “indication” is slightly loose (statins aren’t indicated for NAFLD specifically). The unsupervised split (3×P, HDBSCAN I) shows the methods catching the rare-variant signal (FBLIM1, incidental) rather than the confounding structure. Verdict: expert I more defensible.

8. ramipril × acute myocardial infarction (I21.9) × KLF12 — Expert I vs unsupervised I / I / I / I

A textbook positive control. ACE inhibitors post-MI are a Class I ESC indication (AIRE, HOPE, SAVE — prevention of ventricular remodeling); ramipril is prescribed because of the MI. The curator labels it “direct indication”; KLF12 is incidental.

Expert and all four unsupervised methods agree on Indication — the cleanest result in the set. When the truth is unambiguous confounding-by-indication, the rule cascade and the data-driven clustering converge. Verdict: correct and unanimous — a reassuring validation that the pipeline gets clear cases right.

9. clopidogrel × prosthetic-device complications (T82.8) × HPSE2 — Expert I vs unsupervised I / P / P / I

Dual antiplatelet therapy (clopidogrel + aspirin) after coronary stenting is Class I ESC — clopidogrel is given because the patient has the device, the curator’s “maximum confounding.” FAERS is STRONG (heavy co-reporting), HPSE2 (heparanase-2, urinary/bladder gene) incidental.

The expert Indication call is correct. The unsupervised methods are split (k-means + HDBSCAN agree I; GMM + LCA drift to P) — the GMM/LCA dissent is the rare-variant axis mistaking a confounded indication for predisposition. Verdict: expert right; exposes a real weakness — strong confounding-by-indication can be misread as Predisposition when the variant is rare.

10. ramipril × psoriasis (L40.9) × GLP1R — Expert C vs unsupervised P / P / P / I

The most biologically interesting — both drug and gene have real literature. ACE inhibitors are associated with psoriasis (meta-analysis pooled OR 1.52, 95% CI 1.16–2.00; MR + pharmacovigilance). Independently, GLP1R has genuine psoriasis biology: GLP-1R expression is increased in psoriatic plaques, and genetic proxies of GLP1R expression are associated with psoriasis/psoriatic-arthritis risk. Unlike the other nine genes, GLP1R is not obviously incidental.

Yet the expert labels this Comorbidity/inConclusive and the statistics are thin (FAERS weak, N_INDEPENDENT=0 — no replication). A fascinating tension: the biology is more compelling than the statistics. The expert undercalls it (driven by weak replication); the unsupervised methods scatter without recognizing the GLP1R–psoriasis link, because that biology isn’t in their feature space. Verdict: both classifiers miss the interesting signal — a case where manual, biology-aware review outperforms either automated classifier; deserves follow-up despite weak replication.


Part B — two Effect (E) examples

11. citalopram × osteoporosis (M81.9) × DLEU1 — Expert E vs unsupervised P / P / P / I

This is a genuine drug-induced effect. SSRI-associated bone loss is well documented: a meta-analytic fracture RR ≈ 1.6, decreased bone mineral density, and FDA/EMA warnings. The mechanism is biologically specific — serotonin transporters are present on osteoblasts, osteoclasts and osteocytes; SSRIs impair osteoclast function peripherally but trigger a central serotonin-dependent sympathetic rise that increases bone resorption, with a net effect of bone loss. Notably, the gene DLEU1 is a documented regulator of osteoclastogenesis — so unlike most atlas genes, this one has a plausible bone mechanism, and the curator flags the “tissue rescue” by literature.

The expert Effect call is biologically correct — citalopram causes osteoporosis through an established serotonin–bone pathway. Strikingly, all four unsupervised methods miss it, calling Predisposition (3/4) or Indication (HDBSCAN). This is the concrete manifestation of the structural finding that Effect does not separate as a cluster: even a literature-validated, mechanism-supported drug-induced effect gets absorbed into the rare-variant Predisposition axis. Verdict: expert E correct; unsupervised methods systematically fail to recover Effect.

12. diazepam × COPD (J44.9) × PDE1C — Expert E vs unsupervised P / P / P / I

Benzodiazepines in COPD are a clinically important, well-established hazard: benzodiazepine use is associated with a ~45% increased risk of COPD exacerbations, respiratory failure, and increased mortality, via GABA-A-mediated central respiratory depression (hypoxemia, reduced respiratory drive, hypercapnia). PDE1C (a phosphodiesterase) is expressed in airway/vascular smooth muscle, giving the gene-level signal more plausibility than most. The curator notes the dual reality: benzodiazepines are relatively contraindicated in COPD (respiratory depression), yet heavily used in this anxious/dyspneic population — so both a real ADR and confounded co-prescription are present.

The expert Effect call is defensible — diazepam mechanistically worsens respiratory outcomes in COPD. Again, all four unsupervised methods miss it (P/P/P/I), confirming that the data-driven approach cannot recover the Effect class even when the causality is clinically established. Verdict: expert E correct (with a confounding caveat from co-prescription); unsupervised methods fail to recover Effect, as in example 11.


Honest advice — patterns across all 12

1. Gene-level calls are mostly noise. Of twelve genes, only three have real biology (GLP1R–psoriasis, DLEU1–osteoclastogenesis, PDE1C–airway). The other nine are incidental rare variants. This is the single biggest caveat for the atlas: it treats gene-level GWAS hits as meaningful, but at the individual-association level most are spurious. The real, checkable signal is almost always the drug→ADR epidemiology, not the specific variant.

2. The expert cascade’s Effect tier is largely sound. Two of the three Effect cases examined (SSRI/osteoporosis, benzodiazepine/COPD) are genuine, literature-validated drug-induced effects with plausible mechanisms; only one (statin/varicose) was biologically backwards. The cascade also reads confounding-by-indication correctly (#7, #8, #9). Its main weakness is being conservative on drug-induced ADRs — calling real/contested effects (atenolol & ramipril psoriasis, ramipril disorientation) “Comorbidity.”

3. The unsupervised methods cannot recover Effect, and over-call Predisposition. In both genuine Effect cases (#11, #12) all four methods said P or I, never E. And for confounded indications with rare variants (#3, #5, #9) they over-call P. The rare-variant standard-error axis dominates their behaviour, biasing them toward Predisposition regardless of biology. This is the concrete, literature-grounded version of the T4-recovery failure documented quantitatively elsewhere.

4. Neither classifier is ground truth — use as a screen. Treat the classification (either version) as a prioritization layer, not a verdict. Every association destined for the manuscript needs the manual triage performed here: (a) is the drug→ADR link real, protective, or confounded? (b) is the gene plausible or incidental? The cases worth manual follow-up are the disagreements (#1, #2, #10) and the biology-rich outliers (#10 GLP1R, #11 DLEU1) — not the unanimous ones.


Sources