QSAR Modeling and Fragment Contribution Analysis of N-Benzyltryptamines at Human 5-HT2 Receptor Subtypes

Serotonin 5-HT2 receptor agonists are returning to psychiatric drug development, but separating the three receptor subtypes pharmacologically remains difficult, and agonism at 5-HT2B carries a risk of valvular heart disease. A recent study reported affinities of 46 N-benzyltryptamines and N-benzyl-5-methoxytryptamines at all three human subtypes and analyzed them with an additive substituent model, which gave fragment coefficients of the same sign at every subtype and suggested that selectivity within this scaffold would be hard to achieve. We revisited that conclusion with non-additive models built under a protocol designed for small datasets: fifty stratified partitions with all filtering and hyperparameter tuning inside each training set, external statistics reported as distributions, response permutation on three partitions per endpoint, and partitions that withhold an entire scaffold or an entire substitution position. We repeated the analysis with two descriptor sets that share no variables and are built on different principles. Penalized linear models were predictive and transferable, whereas gradient boosting reproduced the training set exactly, and neither tree ensemble transferred to a held-out scaffold. Fragment contributions at 5-HT2A and 5-HT2C correlated at 0.25 compared with 0.77 for whole-molecule affinities, and both descriptor sets gave the same figure within 0.01. All models failed on para-substituted analogs, which distance-based applicability-domain criteria did not flag, whereas models trained without the ortho or the meta block stayed above a mean-predictor baseline. Aggregation over the whole molecule, rather than descriptor choice, conceals subtype divergence. One group drives the divergence: the N-benzyl substituent raises the predicted 5-HT2A affinity and lowers the predicted 5-HT2C affinity; excluding it raised the correlation to 0.93.