DeepMind’s Improved AlphaFold Model Helping in Drug Discovery

Written By: Aniqa Batool
Last Updated On: November 30, 2023

DeepMind's Improved AlphaFold Model Converts Drug Discovery

DeepMind, one of Google’s AI-centered research teams, released AlphaFold over five years ago, an AI system that can properly predict the shapes of many proteins inside the human body. DeepMind has since enhanced the technique, releasing AlphaFold 2.

DeepMind said today that the latest AlphaFold release can predict virtually all compounds in the Protein Data Bank, the world’s biggest open-access collection of biological molecules.

The capabilities of the new model go beyond protein prediction. DeepMind claims that its algorithm can also predict the molecular makeup of ligands in molecules that bind to “receptor” proteins and shift how cells communicate, nucleic acids molecules that contain key inherited information, and post-translational modifications.

Pharmaceutical researchers are currently using simulations known as “docking methods” to predict how proteins and ligands would interact. A reference protein structure and a recommended place on the structure for the ligand to bind to are required for docking approaches.

However, using the most recent AlphaFold, there is no requirement to use an analogous protein structure or suggested location. DeepMind claims that the model can predict proteins that haven’t been “structurally characterized” before, while also modeling how proteins and nucleic acids communicate- a level of modeling that isn’t achievable with today’s docking approaches.

DeepMind said in a blog that analysis also shows that their model greatly beats AlphaFold on some protein structure prediction issues that are relevant for drug discovery, like antibody binding. Their model’s dramatic leap upward shows the potential of AI to significantly improve their knowledge of the molecular machines that make up the human body. The latest AlphaFold is not without flaws.

DeepMind and Isomorphic Labs researchers reveal in a whitepaper detailing the system’s strengths and limitations that the system falls short of the outstanding method for estimating the structures of RNA molecules — the molecules that carry the instructions for making proteins.