The Psychedelic hERG Data Gap: What 59 Compounds, One Patch-Clamp Anchor, and the Executive Order Mean Together
Pre-IND cardiac safety triage for psychedelic-class compounds under the April 2026 federal psychedelic research order is currently running against a confirmed empty published-literature surface. A primary-source guide to the Paper II pre-registered cross-family hERG QSAR landscape analysis, the cross-architecture prediction divergence it surfaced, and what sponsors, regulatory counsel, grant applicants, and policy readers should plan for.
The federal pipeline now has a deadline. The cardiac safety data does not.
On April 18, 2026, President Trump signed Accelerating Medical Treatments for Serious Mental Illness. The order directs the HHS Secretary to allocate $50 million through the Advanced Research Projects Agency for Health to match state psychedelic-medicine research investment, instructs the FDA to prioritize review of psychedelic compounds, and opens a federal Right to Try access pathway with ibogaine named explicitly in the order text. The signing ceremony featured Joe Rogan, the Luttrell brothers, and W. Bryan Hubbard of Americans for Ibogaine. The political driver is veterans access; the regulatory mechanism is accelerated review.
The order has been read across the regulatory affairs bar as the largest shift in psychedelic-medicine policy since the 2018 Right to Try statute. Within thirty days, DemeRx received an FDA IND acceptance for DMX-1001 (a noribogaine prodrug, accepted 2026-04-23) and Lykos Therapeutics signaled a Type C meeting on MDMA-assisted therapy NDA resubmission following the 2024 Complete Response Letter. atai Life Sciences, Compass Pathways, Cybin, MindMed, Delix Therapeutics, GH Research, Beckley Psytech, and Usona Institute each operate at least one psychedelic-class compound on a clinical pathway that intersects this order. The funding now mobilizing, both federal and venture, is concentrated on compound classes that need pre-IND cardiac safety packages.
A pre-IND cardiac safety package, at a minimum, contains a quantified hERG channel liability assessment. That assessment is generated by one of two paths: experimental electrophysiology (manual patch-clamp on heterologously expressed hERG channels, typically HEK293 or CHO cell lines, run in-house by sponsors or by contract labs including AnaBios, Charles River, ChanTest, Cytocentrics, Eurofins Cerep, Sophion) or computational QSAR prediction (Pred-hERG, ADMET-AI, admetSAR, ADMETlab, OCHEM, Percepta, and others). The accelerated pathway compresses both. Sponsors have less time for experimental campaigns and more reliance on QSAR triage to prioritize which compounds actually need the wet-lab work.
Paper II asks a question the executive order's timeline makes operational: what is the actual state of the published primary hERG patch-clamp literature for the psychedelic-class compounds the federal pipeline is now racing toward? The answer determines whether QSAR triage is doing real work or producing predictions whose uncertainty is not being communicated to downstream decision-makers.
59 compounds, one substantive paired patch-clamp record
Paper II's Phase 1 sourced same-preparation Ki and IC50 paired records across three structurally independent psychedelic compound families. The structural independence matters: tryptamines (indole pharmacophore), phenethylamines (substituted benzene-ethylamine), and cathinones (α-keto-aryl-amine) share no Murcko scaffold and span pharmacologically distinct binding modes. If under-characterization were a one-family literature artifact, it would not survive replication across three structurally and pharmacologically independent families.
It does survive. Across all three families, indexed-database retrieval through ChEMBL (CHEMBL240 hERG target), DrugMatrix (AID 625171 [3H]astemizole displacement), and BindingDB (UniProt Q12809, 724 hERG affinity records, no cutoff) returned exactly one substantive paired patch-clamp datapoint: bupropion, IC50 = 69 µM, a weak blocker at the conventional 10 µM cardiac-safety threshold. All other 58 compounds returned absent, null, or with annotations not meeting the same-preparation Ki + IC50 operational definition.
| Family | N | ChEMBL hERG records (any) | ChEMBL paired Ki+IC50 | BindingDB substantive hits | Out-of-distribution vs Pred-hERG training |
|---|---|---|---|---|---|
| Tryptamines | 22 | 2 (sumatriptan, melatonin; both null) | 0 | 0 of 724 | 21/22 95% |
| Phenethylamines | 20 | 1 (amphetamine, non-patch-clamp) | 0 | 0 of 724 | 19/20 95% |
| Cathinones | 17 | 2 (bupropion, diethylpropion) | 1 (left-censored) | 1 (bupropion 69 µM) | 17/17 100% |
| Total | 59 | 5 | 1 | 1 | 57/59 (96.6%) |
The Phase 1 sourcing was rigorous. Each compound carried a PubChem CID, canonical isomeric SMILES, InChIKey, and molecular formula audit; per-family rosters are deposited in supplementary information at the OSF project. The 59-compound scope deliberately spanned regulatory-pipeline status (FDA-approved bupropion and diethylpropion; the seven marketed triptan migraine therapeutics; MAPS-pipeline MDMA; COMPASS and Usona-pipeline psilocybin), classical research-chemical pharmacology (DMT, 5-MeO-DMT, mescaline, the 2C-x and DOx series), bath-salt potent-blocker candidates (MDPV, α-PVP, methcathinone, naphyrone), and endogenous comparators (melatonin, β-phenethylamine). Under-characterization is not specific to one regulatory tier.
The pre-registration anticipated this possibility but assigned it low prior probability. Six of eleven pre-execution priors registered with explicit probability and falsification disposition were falsified by the empirical state. The falsifications cluster on a single meta-assumption: "well-known compound classes, marketed pharmaceuticals, and regulatory-pipeline compounds have published primary patch-clamp data with same-preparation Ki + IC50 pairing." The implicit mental model that drives QSAR-triage trust is systematically miscalibrated for psychedelic-class compounds. The pre-registration discipline is what made this finding visible rather than letting it fail quietly during sourcing.
The ergoline boundary observation produced a substantive secondary finding. PMID 14660002 (Hurst 2003, Eur J Pharmacol) reports primary hERG patch-clamp on pergolide with quantified IC50 = 0.12 µM, a potent blocker, but unpaired (no same-preparation Ki measurement). Pergolide was withdrawn from the US market in 2007 for 5-HT2B-mediated cardiac valvulopathy. Even pharmaceuticals withdrawn for cardiac safety reasons lack same-preparation Ki + IC50 records at the published-literature level. The paired-record gap is structural to publication conventions across pharmaceutical chemistry, not specific to under-studied compound classes. This sharpens the operational interpretation: the gap will not close by accident.
Which compounds, specifically
Sponsors and regulatory affairs counsel reading this article often need to know whether a specific compound on their development roadmap is in scope. The 59-compound Paper II Phase 1 set is enumerated in the OSF deposit; the named composition by family is summarized below.
What this does to QSAR-based pre-IND triage
QSAR triage is the standard fallback when experimental electrophysiology is not yet available, and it is the standard accelerator inside experimental campaigns to prioritize which compounds get bench time first. Three production hERG QSAR systems are used widely in pre-IND submissions and academic safety work: Pred-hERG 5.0 (Sanches, Braga, Alves, Andrade 2024, LabMol UFG; LightGBM gradient-boosted classifier on ECFP4 fingerprints with a multiclass + regression consensus rule), ADMET-AI 2.0.1 (Swanson et al. 2024, Stanford; Chemprop graph neural network), and admetSAR 2.0 (Yang et al. 2019, ECUST; support vector machine on hand-crafted descriptors). Paper II ran all three architectures against all 59 compounds.
The within-family positive-call rates diverge by up to an order of magnitude. The cathinone family produces the cleanest illustration: ADMET-AI calls 10 of 17 cathinones (59%) Blocker; Pred-hERG consensus calls 1 of 17 (6%) Blocker. The same compound input, three production architectures, a 10× rate spread.
| Family | ADMET-AI GNN | admetSAR SVM | Pred-hERG consensus | Spread (max-min) | Ratio (max/min) | S2.1 threshold |
|---|---|---|---|---|---|---|
| Tryptamines (N=22) | 82% | 77% | 36% | 45 pts | 2.25× | CLEARED |
| Phenethylamines (N=20) | 80% | 55% | 65% | 25 pts | 1.45× | CLEARED via SD |
| Cathinones (N=17) | 59% | 18% | 6% | 53 pts | 10.00× | CLEARED |
The pre-registered effect-size threshold (the larger of 2× positive-call rate ratio or 1 standard deviation of pooled within-family bootstrap iterates; pre-commit S2.1, locked at OSF publish 2026-05-13) cleared on three of three families. The cross-architecture positive-call divergence is robust, not marginal. Critically, the pre-committed threshold sits before the data was collected, and the per-family Wilson 95% confidence intervals are honestly reported (cathinones' non-overlap is clean on the largest architecture pair; phenethylamines' CIs overlap pairwise, indicating real magnitude uncertainty even at confirmed-divergence-level effect size).
The interpretation requires a load-bearing caveat. Cross-architecture positive-call divergence measures cross-architecture decision-behavior on a fixed input set; under the data-landscape constraint, it cannot measure deviation from per-compound experimental ground truth because that ground truth is empirically absent for 58 of 59 compounds. The single-anchor case (bupropion) is the one ground-truth-tied validation: against the experimental Non-blocker classification at IC50 = 69 µM, ADMET-AI calls Blocker at 0.609 probability (wrong, near decision boundary), admetSAR calls Non-blocker at 0.726 (correct), Pred-hERG binary calls Non-blocker at 86.9% (correct), Pred-hERG regression over-estimates potency by approximately 16×, and Pred-hERG consensus correctly outputs Non-blocker because the rule weights binary appropriately in this direction. One anchor, three architectures, two different outcomes against ground truth.
Why scaffold-in-training is not the same as in-distribution
The applicability domain (AD) flagging mechanism is the standard way a QSAR system warns downstream users that a query compound sits outside the supported training region. Pred-hERG's AD flagging treats Murcko scaffold membership as evidence of training coverage. Paper II shows why that heuristic is insufficient for the psychedelic-class case.
21 of the 59 compounds have a Murcko scaffold that appears in the Pred-hERG curated training set (3,918 unique scaffolds across 7,307 training compounds). Twenty of those 21 also have a max-Tanimoto ECFP4 bit-level distance to the training set greater than 0.50, the conventional out-of-distribution cutoff. The phenethylamine and cathinone Murcko cores are common in the Pred-hERG training distribution because the cores are shared with dopamine reuptake inhibitor and β-keto-amphetamine training compounds; the substitution patterns that distinguish psychedelic-class compounds from those training compounds are bit-level distant.
The pre-registered bidirectional hypothesis (pre-commit S3.2) anticipated two effects on this subset: fingerprint architectures (Pred-hERG) would fail worst on scaffold-common-but-bit-level-distant compounds because the AD flagging mechanism would not flag, and scaffold-aware architectures (ADMET-AI GNN, admetSAR SVM) would perform better because scaffold context would be correctly recognized. The empirical outcome was partial confirmation in the scaffold-aware direction only. Combined ADMET-AI plus admetSAR positive-call rate on the scaffold-in-training-bit-distant subset is 47.8% of the rate on the scaffold-novel-bit-distant subset, clearing the ≤ 0.5 threshold and demonstrating that graph-encoding architectures do preferentially capture scaffold context relative to bit-level features.
The Pred-hERG worst-case half failed in the opposite direction of prediction: Pred-hERG does not behave anomalously on scaffold-recognized compounds. It simply maintains a lower overall positive-call rate.
Who this hits, named
The compound classes Paper II surveys are not academic curiosities. They are the active substrate of the post-EO regulatory and venture pipeline.
| Sponsor / Program | Compound class | Stage as of May 2026 | Paper II scope |
|---|---|---|---|
| DemeRx Inc. | Noribogaine (DMX-1001) | FDA IND accepted 2026-04-23 (Phase 2 alcohol use disorder) | Paper I anchor; cross-family extension via psychoactive tryptamines |
| Lykos Therapeutics (formerly MAPS PBC) | MDMA-assisted therapy | 2024 CRL; Type C meeting / NDA resubmission path under EO | Phenethylamine family Tier C; zero published paired records |
| COMPASS Pathways | Psilocybin (COMP360) | Phase 3 TRD; Phase 3 PTSD enrolling | Tryptamine family Tier B; zero published paired records |
| Cybin Inc. | Deuterated psilocin analogs (CYB003) | Phase 3 MDD | Tryptamine family analog space; zero published paired records |
| MindMed | LSD analog (MM120) | Phase 3 GAD; Phase 3 MDD | Lysergamide / ergoline boundary; zero paired records on parent class |
| atai Life Sciences (Recognify) | R-MDMA, ibogaine analogs | Phase 2 portfolio | Both phenethylamine and tryptamine families in scope |
| Delix Therapeutics | Non-hallucinogenic psychoplastogens (tabernanthalog lineage) | Phase 1 | Paper I scaffold-aware over-call pattern documented on TBG; same architecture-failure family |
| GH Research | Mebufotenin (5-MeO-DMT) inhalation (GH001) | Phase 2/3 TRD | Tryptamine family Tier B; zero published paired records |
| Usona Institute | Psilocybin (PSIL201) | Phase 2/3 MDD | Tryptamine family Tier B; zero published paired records |
| Beckley Psytech | 5-MeO-DMT (BPL-003) | Phase 2b TRD | Tryptamine family Tier B; zero published paired records |
The pattern across the sponsor table is uniform. Every program above is advancing a compound or compound class for which the published primary hERG patch-clamp literature does not contain a same-preparation Ki + IC50 paired record. Cardiac safety packages will be generated, but they will be generated by experimental campaigns the sponsor runs (in-house or via contract labs) rather than by literature aggregation, because the literature does not supply the data. The pipeline is not blocked; it is simply working against a data-infrastructure gap that the published literature alone cannot bridge.
What sponsors, regulatory counsel, grant applicants, and funders should plan for
The Paper II Discussion enumerates four audience-specific recommendation tracks. The condensed operational guide is below.
For sponsors filing pre-IND cardiac safety packages
Three concrete moves. First, document the training-distribution membership status of every QSAR-predicted compound: Murcko scaffold inclusion test against the published curated training set, and max-Tanimoto ECFP4 bit-level distance. Both are one-line RDKit calls and both materially affect how reviewers should weight the prediction. Second, report per-sub-model output (binary classifier, multiclass, regression) where consensus aggregation rules are used. The Pred-hERG noribogaine and Paper II psilocin cases show that consensus aggregation can override correct binary calls. Per-sub-model transparency prevents silent false-clears. Third, budget for experimental electrophysiology campaigns (manual patch-clamp on heterologously expressed hERG via vendors including AnaBios, Charles River Laboratories Discovery, ChanTest by Charles River, Cytocentrics, Eurofins Cerep, Sophion Bioscience) early rather than late. The published-literature surface will not supply the missing data within any timeline the executive order's accelerated review permits.
For regulatory affairs counsel
The cardiac safety question under accelerated review is operational, not abstract. The pre-IND submissions sponsors file now will be reviewed by FDA staff who are themselves operating under accelerated-pathway pressure. Counsel work that adds value: (a) reviewing QSAR-prediction documentation in submissions for training-distribution-membership reporting and per-sub-model transparency; (b) flagging compounds with scaffold-in-training-but-bit-distant status as warranting explicit AD discussion in the cardiac safety section; (c) coordinating with sponsors on experimental-campaign timing so wet-lab work can begin before regulator-meeting feedback identifies the gap. Independent-counsel commentary from Frier Levitt and the Petrie-Flom Center at Harvard Law has framed the cardiac data gap as a defining clinical risk of the order's expansive ibogaine pathway; the Paper II finding extends the risk frame cross-family.
For grant applicants targeting the data infrastructure gap
The data-landscape finding identifies one concrete fundable gap. Coordinated experimental campaigns generating same-preparation Ki + IC50 paired records on the 59 Phase 1 compounds, or a representative subset prioritized by regulatory urgency, would directly address it. Priority candidates: bath-salt potent-blocker candidates (MDPV, α-PVP, naphyrone) where cardiotoxicity signal exists in forensic toxicology and emergency medicine but not in patch-clamp; FDA-pipeline psychedelic-class compounds (MDMA via Lykos pipeline; psilocybin via COMPASS / Usona / Cybin / Beckley pipelines; ibogaine via DemeRx DMX-1001 pipeline) where cardiac safety packages are required for IND advancement; the 20 scaffold-in-training-but-bit-distant compounds where architecture-specific failure modes can be directly resolved by experimental measurement. Funding mechanism candidates: NIDA (the National Institute on Drug Abuse maintains active interest in cardiotoxicity of psychoactive compounds); NIMH (consistent with the EO's serious-mental-illness framing); FDA Broad Agency Announcement mechanism; ARPA-H psychedelic-medicine portfolio under the $50M EO match-funding stream; private foundations including the Steven and Alexandra Cohen Foundation, Source Research Foundation, and the Riverstyx Foundation. Grant proposals citing Paper II's empirical state directly anchor the gap-quantification claim in a pre-registered, peer-public analysis.
For QSAR / chemoinformatics developers
The three production architectures evaluated in Paper II (Pred-hERG, ADMET-AI, admetSAR) each receive specific recommendations in the paper. In summary: per-sub-model output transparency in deployed service interfaces (all three teams); joint (scaffold-in-training, bit-level-distance) AD reporting (Pred-hERG / LabMol UFG specifically); probability-calibration documentation on scaffold-aware compound classes (ADMET-AI / Stanford); caffeine-reference recalibration and descriptor-space audit on psychedelic-class compounds (admetSAR / ECUST). Training-set augmentation requires upstream experimental campaign work first because the published literature does not currently supply the data to augment from.
Where this leaves patients tracking the pipeline
Patients reading this article often arrived through the federal Right to Try pathway or through a treatment program operating under state-level psychedelic-medicine frameworks (Texas, Kentucky, Oregon, Colorado). The Paper II finding does not change what patients should ask treatment providers about cardiac safety. The companion clinical decision guide at omnirx.org/articles/ibogaine-and-the-heart-clinical-decision-guide contains the operational 16-question framework for ibogaine specifically, including baseline 12-lead ECG protocol, electrolyte correction thresholds, continuous telemetry duration, emergency response capability, and medication exclusion list questions.
What the Paper II finding does change is the answer to "how confident is the prediction that this compound is cardiac-safe?" for compounds where the only safety claim rests on QSAR prediction without underlying patch-clamp ground truth. For most psychedelic-class compounds in active pipelines, that confidence interval is wider than the prediction outputs alone communicate. The federal Right to Try pathway specifically allows access to investigational drugs that have completed Phase 1 safety trials; Phase 1 trials on these compound classes will generate cardiac safety data that will substantially update what is currently known.
Patients considering Right to Try access for a specific compound should ask whether the cardiac safety package backing the IND includes patch-clamp data on the specific compound, or whether it relies on QSAR prediction plus related-compound bridging. The honest answer for most psychedelic-class compounds today is that it relies on QSAR prediction plus bridging.
Frequently asked questions
What is the central finding of the Paper II hERG data-landscape analysis?
Why does this matter for the April 2026 federal psychedelic research executive order?
Do QSAR models close the gap?
What does the consensus-rule failure mode mean for sponsors?
Which compound classes are most affected?
What should regulatory affairs counsel and sponsors plan for under accelerated review?
How does this relate to Paper I and to the ibogaine cardiac safety question?
Is Paper II peer-reviewed?
Source ledger
Every empirical, regulatory, and policy claim in this guide rests on a primary source. The Paper II pre-registration and Paper I preprint are cited only for claims about the data-landscape finding, the cross-architecture positive-call divergence, and the consensus-rule failure mode documented in those papers. All other claims rest on regulatory, peer-reviewed, and government sources listed below.
Corrections policy. If you identify a factual error, an outdated citation, or a primary-source disagreement with anything stated here, contact vinnycouey@gmail.com. Substantive corrections are logged in the corresponding OSF audit-trail entry with an explanatory note appended to this article's revision history. Paper II preprint coordinates (ChemRxiv DOI) will be backfilled here and into the JSON-LD citation graph once the ChemRxiv submission resolves.