Systematic Under‐Characterization of Psychedelic Compound Classes in Published hERG QSAR Training Data: A Pre‐Registered Multi‐Family Analysis

Abstract

QSAR models for hERG cardiotoxicity are increasingly deployed in pre‑IND cardiac safety triage of psychedelic‑class therapeutics under accelerated regulatory review. We make two contributions. Empirical: across three structurally independent psychedelic compound families (22 tryptamines, 20 phenethylamines, 17 cathinones; 59 compounds spanning natural products to FDA‑approved medications), published primary hERG patch‑clamp literature with same‑preparation Ki + IC50 pairing yields one substantive datapoint (bupropion, IC50 = 69 µM). Six of 11 pre‑registered priors that anticipated marketed‑pharmaceutical compounds would have published patch‑clamp data were falsified in a tightly clustered pattern on the assumption “well‑known compound classes have published data.” A pre‑registered targeted full‑text retrieval protocol (PubMed bath‑salt toxicology, Drugs@FDA bupropion NDA filings, ergoline boundary spot‑check) surfaced zero additional pairs, resolving the central thesis to CONFIRMED at published‑literature level. Three‑architecture Sub‑analysis 2 on all 59 compounds (Pred‑hERG 5.0, ADMET‑AI 2.0.1, admetSAR 2.0) cleared the pre‑registered cross‑architecture positive‑call divergence (CAPD) threshold (effect‑size criterion S2.1) on three of three families at the surviving data scale under the (C) data‑landscape constraint; cathinones show 10× spread between ADMET‑AI Blocker rate (59%) and Pred‑hERG consensus (6%). Two suggestive (N = 2) cross‑family cases of the Paper I Pred‑hERG consensus‑rule override are observed (psilocin and a 4‑hydroxy‑N‑methyl‑N‑ethyl tryptamine analog: binary correctly calls Blocker, consensus overrides to Non‑blocker); systematic characterization of the asymmetric‑failure‑mode hypothesis requires larger‑scale evaluation (§4.5). The pre‑registered bidirectional scaffold‑vs‑bit‑level hypothesis resolved PARTIAL CONFIRMATION on the scaffold‑aware half: combined ADMET‑AI + admetSAR positive‑call rate on scaffold‑in‑training‑but‑bit‑level‑distant compounds is 47.8% of the rate on scaffold‑novel‑bit‑distant compounds, confirming scaffold‑context recognition by graph‑encoded and descriptor‑encoded architectures. Methodological: we adapt the PRISMA‑S search‑protocol‑asload‑bearing‑methodology pattern (Rethlefsen et al. 2021) to chemoinformatics QSAR‑validity work and offer the data‑landscape‑primary template as a worked example for chemoinformatics adopters; the template pre‑registers a search protocol as load‑bearing methodology, ties central‑thesis disposition to retrieval outcome, and frames secondary contributions as evidence‑for the primary finding. The empirical finding implies pre‑IND cardiac safety triage of psychedelic‑class compounds operates against a data infrastructure that QSAR architectures cannot bridge from training to query distribution; the methodological contribution generalizes as a reusable template for meta‑empirical literature‑state claims in QSAR‑validity work.