StereoAtlas: A Large-Scale Dataset of Chiral Compounds with Distinct Bioactivities Curated from ChEMBL37
Year: 2026
Type: preprint
Abstract: Stereoisomers share identical two-dimensional molecular topology and exhibit a Tanimoto similarity score of exactly 1.0 under standard fingerprint methods, rendering their bioactivity differences invisible to conventional data-driven or physics-based activity prediction approaches. Here, we present StereoAtlas, a dataset of 155,526 experimentally measured true isomer families derived from ChEMBL37, in which paired or families of stereoisomers exhibited pharmacological activity for the same biological target, in the same organism. Fold differences in activity between stereoforms were computed under two complementary strategies — MinMax and reference-form— and classified into five stereoselectivity tiers: None (<2×), Mild (2–5×), Moderate (5–10×), Strong (10–50×), and Extreme (>50×). Among 38,446 IC50 families, 9.5% exceed the 50-fold Extreme threshold, and the maximum fold difference observed is 195,000,000× at the oxytocin receptor (Kd) for the stereoisomeric families CHEMBL386547 (potent isomer) and its inactive counterparts (CHEMBL274635 and CHEMBL414373). Spanning 420,721 unique compounds, 4,519 biological targets, 471 organisms, and 879 activity types, this dataset provides a detailed, systematic, large-scale experimental resource for chirality cliff-based analysis. All data and pipeline scripts are publicly available.
Source: ChemRxiv
Language: English
Subfield: Computational Theory and Mathematics
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7220476484