Almost half a decade ago, DeepMind, a prominent AI research lab under Google’s umbrella, introduced AlphaFold, an AI system renowned for its precise predictions of the structures of numerous proteins within the human body. Since then, DeepMind has continued to refine AlphaFold, unveiling an enhanced iteration named AlphaFold 2 in 2020.
The work at DeepMind persists, and today they unveiled the latest iteration of AlphaFold, the successor to AlphaFold 2. This cutting-edge AlphaFold can now generate predictions for nearly all molecules present in the Protein Data Bank, which happens to be the largest open-access repository of biological molecules in the world.
Furthermore, Isomorphic Labs, a spin-off company affiliated with DeepMind and specializing in drug discovery, is actively leveraging the capabilities of this new AlphaFold model, which it co-developed. Isomorphic Labs is employing this AI advancement for therapeutic drug design, aiding in the characterization of various types of molecular structures crucial for disease treatment.
The latest AlphaFold is not limited to protein structure prediction. DeepMind asserts that this model can accurately predict the structures of ligands—molecules that bind to “receptor” proteins and trigger changes in cellular communication. It can also predict the structures of nucleic acids (molecules containing vital genetic information) and post-translational modifications (chemical alterations occurring after a protein’s formation).
The prediction of protein-ligand structures is a valuable asset in drug discovery, as it facilitates the identification and design of novel molecules with potential for drug development. Presently, pharmaceutical researchers employ computational simulations known as “docking methods” to assess the interaction between proteins and ligands. These methods require specifying a reference protein structure and a suggested binding position on that structure for the ligand.
However, with the latest AlphaFold, there’s no longer a need for a reference protein structure or a suggested position. The model can predict the structures of proteins that haven’t been structurally characterized before. It can also simulate how proteins and nucleic acids interact with other molecules, a level of modeling that surpasses what can be achieved with today’s docking methods.
DeepMind states, “Early analysis also shows that our model greatly outperforms [the previous generation of] AlphaFold on some protein structure prediction problems that are relevant for drug discovery, like antibody binding.” This demonstrates the potential of AI to significantly enhance our scientific understanding of the molecular machinery within the human body.
Nevertheless, the newest AlphaFold is not flawless. In a whitepaper that outlines the system’s strengths and limitations, researchers from DeepMind and Isomorphic Labs acknowledge that the system falls short of the most advanced method for predicting the structures of RNA molecules, which play a crucial role in encoding protein instructions. Undoubtedly, both DeepMind and Isomorphic Labs are actively working to address this limitation.
