DeepMind's AlphaFold predicts psychedelics
- Science and Nature coverage point to a narrower story: AlphaFold helped researchers prospectively find new ligands for serotonin and TAAR1 receptors linked to psychedelics. - In Science, teams docked 1.6 billion molecules to an AlphaFold2 5-HT2A model and found hit rates comparable to experimental structures. - The work shifts AlphaFold from structure prediction toward drug screening for psychiatric targets. (nature.com)
Proteins are folded chains that act like tiny locks in the body, and drugs work by fitting those locks. AlphaFold is DeepMind’s artificial-intelligence system for predicting a protein’s 3D shape from its amino-acid sequence. (deepmind.google) That matters in drug discovery because many useful targets do not have lab-measured structures from X-ray or cryo-electron microscopy. A predicted shape can give chemists a starting map for testing which molecules might bind. (deepmind.google) (science.org) The psychedelic angle comes from receptors such as serotonin 2A, or 5-HT2A, which is a main target for LSD and psilocybin-like effects. Another target, trace amine-associated receptor 1, or TAAR1, is being studied for schizophrenia and other neuropsychiatric disorders. (science.org 1) (science.org 2) One key paper was published in Science in 2024 by Jiankun Lyu, Bryan Roth, Brian Shoichet and colleagues. They used AlphaFold2 models to guide prospective ligand discovery for sigma-2 and 5-HT2A receptors before comparing the results with experimental structures. (science.org) For 5-HT2A, the team docked 1.6 billion molecules against both an AlphaFold2 model and an experimental structure. After synthesizing and testing more than 100 top-ranked compounds for each setup, they reported similar hit rates and hit affinities from the model and the lab structure. (science.org) The paper also said the cryo-electron microscopy structure of one agonist, Z7757, matched the docking pose predicted from the AlphaFold2 model. That result addressed a central criticism of structure prediction in drug discovery: whether a model can place a small molecule in the right pocket, in the right orientation. (science.org) A second 2024 study in Science Advances pushed the same idea at TAAR1, a receptor without an experimental structure at the time. Alejandro Díaz-Holguín, Jens Carlsson, Per Svenningsson and colleagues docked more than 16 million compounds to AlphaFold and homology models of TAAR1. (science.org) They experimentally tested 30 compounds from the AlphaFold screen and 32 from the homology-model screen. The paper reported 25 TAAR1 agonists overall, with the AlphaFold screen delivering a 60% hit rate, more than twice the homology model’s rate, and finding the most potent agonists at 0.03 micromolar. (science.org) Nature’s January 18, 2024 news article framed those papers as evidence that AlphaFold can be useful in drug discovery when researchers pair the models with massive virtual screens and then verify the hits in the lab. The article described researchers finding “hundreds of thousands” of potential psychedelic-like molecules in one screening effort. (nature.com) So the clean version of this story is not that DeepMind published a new psychedelic prediction system this week. It is that AlphaFold models have already been used to search for new compounds at receptors tied to psychedelic effects and psychiatric drug development, with lab follow-up showing the models can sometimes work as well as experimental structures. (science.org 1) (science.org 2) (nature.com) The next step is the slow part: synthesis, receptor assays, animal studies and, if compounds hold up, human trials. AlphaFold can narrow the search, but it does not show by itself whether a molecule will be safe, nonhallucinogenic or clinically useful. (science.org 1) (science.org 2)
