AI Tool Unlocks Structural Secrets of Human Bitter Taste Receptors

Researchers have successfully used AlphaFold3, the latest evolution of DeepMind’s AI-powered protein prediction system, to generate accurate three-dimensional structures for all 25 known human bitter taste receptors—collectively known as T2Rs.

Published in Current Research in Food Science, the study underscores AlphaFold3’s capacity to predict previously inaccessible structural information about this enigmatic family of G protein-coupled receptors (GPCRs), which are not only responsible for detecting bitterness on the tongue but also influence key metabolic processes throughout the body.

T2Rs play a pivotal role in the gut-brain axis, contributing to appetite regulation and glucose metabolism. Though these receptors were traditionally studied in the context of taste perception, emerging evidence shows they are also expressed in the gastrointestinal tract and other organs. When activated by bitter compounds—including polyphenols and pharmaceutical agents—T2Rs can trigger the release of hormones and neurotransmitters that improve glucose tolerance and suppress appetite.

Despite their biological importance, only two T2R structures—T2R14 and T2R46—have been experimentally resolved using cryo-electron microscopy (cryo-EM). Crystallizing GPCRs is notoriously difficult, often requiring stabilizing ligands and specific conformations. To address this gap, the research team turned to AlphaFold3, released in beta in 2024, which uses a diffusion model to generate atomic-level predictions with high fidelity.

The study compared AlphaFold3 (AF3) to its predecessor, AlphaFold2 (AF2), across multiple parameters. While AF3 surprisingly yielded slightly lower overall confidence scores (pLDDT) than AF2, it outperformed in structural accuracy when benchmarked against known cryo-EM data for T2R14 and T2R46. Notably, the predicted structure of T2R46 in its active, strychnine-bound state aligned with the experimental model at an impressively low root-mean-square deviation (RMSD) of 1.709 Å, a key indicator of structural similarity.

AF3’s accuracy was particularly strong in the transmembrane (TM) regions of T2Rs, which are structurally conserved across the family. In contrast, greater variability was found in the extracellular loops, especially ECL2, which is believed to be involved in ligand recognition. This divergence likely reflects the loop’s inherent flexibility and diverse ligand-binding functions. The intracellular domains, by comparison, showed high structural consistency, reinforcing the hypothesis that T2Rs share a conserved mechanism for binding the G protein α-gustducin, a signaling partner crucial to their metabolic effects.

The team also leveraged AF3’s capacity to predict protein-protein complexes to model the interaction between T2Rs and G proteins. These predicted complexes closely mirrored known cryo-EM structures, with RMSD values below 2 Å, suggesting that AF3 could serve as a valuable tool for simulating dynamic receptor activation processes and guiding drug design.

Beyond structural prediction, the study explored the sequence and functional diversity of the T2R family. Using clustering algorithms, researchers identified subgroups of receptors with high sequence identity and structural similarity. For instance, T2R19, T2R20, T2R31, T2R43, T2R45, T2R46, and T2R50—all encoded within a ~400 kb region on chromosome 12—clustered tightly, hinting at shared physiological roles. Conversely, T2R14, which is known to possess both extracellular and intracellular ligand-binding sites, clustered with T2R10 and T2R13, suggesting these receptors may also exhibit complex ligand interactions.