Google DeepMind has introduced a new AI model called AlphaProteo, designed to generate novel proteins that bind effectively to target proteins. This breakthrough technology can significantly reduce the time required for initial experiments in biology and health research.
Proteins interact with each other through binding, which is crucial for various cellular processes. While DeepMind’s protein structure prediction tools like AlphaFold have provided valuable insights into how proteins interact, they haven’t been able to design new proteins that bind to them. AlphaProteo changes this by enabling the design of high-strength protein binders.
The AI system can speed up research in various fields, including drug development, cell and tissue visualization, disease understanding and diagnosis, and crop pest resistance. To demonstrate its capabilities, the DeepMind team targeted seven proteins and used AlphaProteo to generate protein binders. The results showed that the candidates were more likely to bind successfully to their targets than previous methods.
AlphaProteo was trained using vast amounts of protein data from the Protein Data Bank and over 100 million structure predictions from AlphaFold. This training enables the AI system to learn how molecules can bind to each other, allowing it to generate candidate proteins that bind to a target protein at its preferred binding site.
The graph below shows the probability of successful binding for proteins generated by AlphaProteo and conventional methods. The results indicate that AlphaProteo has a higher probability of successful binding than conventional methods, requiring fewer candidates to find a protein that can bind successfully.
The figure below displays the affinity score, which measures the strength of binding. The graph is on a logarithmic scale, with smaller numbers indicating stronger binding. AlphaProteo was able to create stronger bonds with all proteins tested.
These results demonstrate that AlphaProteo can significantly reduce the time required for initial experiments. However, it is not possible to design binders for all proteins. Even with AlphaProteo, designing a binder for TNFÉ‘, a protein associated with autoimmune diseases like rheumatoid arthritis, proved challenging.
DeepMind plans to continue improving and extending AlphaProteo to eventually address challenging targets like TNFÉ‘.
Source: https://gigazine.net/gsc_news/en/20240906-google-deepmind-alphaproteo/