Explainable Artificial Intelligence (xAI) for 5-HT2A Receptor Binding Affinity of New Psychoactive Substances.
Verena Schöning, Katharina Elisabeth Grafinger, Daniel Pasin, Christophe P Stove, Wolfgang Weinmann, Felix Hammann
Molecules (Basel, Switzerland) June 1, 2026 DOI: 10.3390/molecules31111888 (opens in new tab) via PubMed
Summary
AI-generated from the abstractNew psychoactive substances (NPS) mimic existing drugs and can be highly potent, leading to frequent intoxications. A key target for hallucinogenic NPS is the 5-HT2A receptor. Measuring receptor binding affinity (Ki) through in vitro assays is resource-intensive. Using publicly available Ki data for 5-HT2A, five classification machine learning models were trained with molecular descriptors and fingerprints. The models achieved precisions and recalls up to 93% and 92%, respectively. Explainable artificial intelligence (SHAP values and similarity maps) was used for interpretation. The results align with previous experiments and support the models' suitability for predicting binding affinities of potential 5-HT2A ligands.
Study at a glance
| Characteristics | Classification model development and validation Peer reviewed |
|---|---|
| Keywords | 5-ht2a Nps Binding affinity Classification model Explainable artificial intelligence |
| Key finding | Five machine learning models trained on publicly available 5-HT2A Ki data achieved precisions and recalls up to 93% and 92%, supporting their suitability for predicting binding affinities of potential 5-HT2A ligands. |
Abstract
New psychoactive substances (NPS) are a heterogeneous group of recreational drugs that mimic the actions and psychoactive effects of existing pharmaceutical products or recreational drugs. As NPS can be highly potent, even exceeding their template compound's potency, there are frequent reports of non-fatal and fatal intoxications. One principal target for hallucinogenic and psychedelic drugs, including NPS, is the 5-hydroxytryptamine receptor 2A (5-HT2A). Since NPS are designed to evade legal restrictions, this drug market is quickly evolving, and researchers are playing catch-up, investigating these novel compounds for their toxicological and pharmacological properties. Receptor binding affinity (Ki) is an important property describing ligand-receptor interactions and is a prerequisite for receptor activation. Competitive in vitro assays can be used to assess Ki; this is a resource-intensive process. We used publicly available Ki data for the 5-HT2A, calculated molecular descriptors and fingerprints, and trained five classification machine learning models. The predictive performance of the models had precisions and recalls up to 93% and 92%, respectively. We used explainable artificial intelligence, i.e., SHAP values and similarity maps, for model interpretation. The results are in line with previous experiments and support its suitability to predict the binding affinities of possible 5-HT2A ligands.