Comment on Suresh et al. (2026): Drug-level sample size, validation and embedding confounds in receptor–semantic mappings

Suresh et al. (2026, Psychopharmacology) related GloVe embeddings of 4,496 first-hand reports of 21 psychedelic drugs to receptor affinity profiles using partial least squares regression (PLSR), and reported associations between specific receptors and experiential themes. Because affinities vary only between drugs, every report of a drug carries the same profile. I show algebraically that within-drug variation drops out of every cross-product with the affinities. The published loadings can be reproduced from the drug means alone (r ≥ 0.998), so the analysis rests on 21 drugs, with a Kish effective sample size of about 11, rather than on 4,496 reports. The model is near-saturated at the drug level. The cross-validation, receptor subset selection and loading interpretation have not been assessed at the drug level, and the associations have not been tested against chance or for generalisation beyond the few heavily weighted drugs, within the sample or to new ones. A reanalysis of published values shows that the first three components closely follow the DMT–MDMA contrast (r = 0.93), mean report length (r = 0.96) and the LSD–DMT contrast (r = 0.97). The length component follows from summing word vectors. I propose drug-level validation, permutation and resampling analyses that can be run on the existing data. Code reproducing all numbers is publicly available.