Where do the ligands bind? Co-folding bitter taste GPCRs with the BitterDB chemical space

Bitterness is a key taste modality mediated in vertebrates by TAS2R G-protein-coupled receptors, which also function in diverse extraoral tissues. Recent cryo-EM structures have revealed a non-classical intracellular pocket in TAS2R14, raising the question of whether ligand pocket choice can be predicted computationally and what sequence features control it. Here we evaluate the Boltz-2 co-folding framework on all currently available agonist-TAS2R cryo-EM complexes and show that it correctly identifies the experimentally observed binding pocket for 12 of 15 pairs, including intracellular binding that docking into predicted receptor models fails to reproduce. Focusing on aristolochic acid, which binds intracellularly to TAS2R14 and extracellularly to TAS2R43, we use a series of in silico morphing experiments to pinpoint transmembrane helices 3 and 7, and specific residues within them, as key determinants of pocket preference. Extending the analysis to [~]1,500 agonist-receptor associations from BitterDB, we find that while most receptors are predicted to bind agonists predominantly in the extracellular pocket, several TAS2Rs may have both extracellularly and intracellularly binding ligands. Finally, by fine-tuning the Boltz-2 affinity module on [~]7,000 positive and negative experimental data points, we obtain a TAS2R-specific classifier that improves AUROC from 0.54 to 0.82 and average precision from 0.24 to 0.58 on a validation set.