MIT researchers are using artificial intelligence (AI) to design new proteins that go beyond those found in nature. They developed machine-learning algorithms that can generate proteins with specific structural features, which could be used to make materials with desired mechanical properties. This breakthrough could potentially replace materials made from petroleum or ceramics, resulting in a smaller carbon footprint.
The researchers employed a generative model, similar to the one used in AI systems like DALL-E 2. However, instead of generating images from natural language prompts, they adapted the model to predict amino acid sequences of proteins with specific structural objectives. The team demonstrated in a paper published in Chem how these models can generate realistic and novel proteins.
Senior author Markus Buehler, a professor in Engineering and professor of civil and environmental engineering and of mechanical engineering, explained that these models can generate millions of proteins in a short period of time. This gives scientists a wide array of new ideas to explore. For example, the tool could be used to develop protein-inspired food coatings, which would keep produce fresh for longer periods while being safe for human consumption.
The researchers used attention-based diffusion models, a new advent in machine learning, to predict amino acid sequences that meet design targets. These models are effective at generating high-quality, realistic data that can meet set objectives. By combining different amino acid structures, the researchers can create proteins with different mechanical properties. For instance, proteins with alpha helix structures yield stretchy materials, while those with beta sheet structures yield rigid materials. Combining both can create stretchy and strong materials, such as silks.
The team compared the new proteins to known proteins with similar structural properties and found some overlap with existing amino acid sequences, suggesting that many of the generated proteins could be synthesized. The researchers also ensured that the predicted proteins were realistic by attempting to trick the models with physically impossible design targets. However, instead of producing improbable proteins, the models generated the closest synthesizable solution.
Going forward, the researchers plan to experimentally validate some of the new protein designs in a lab. They also aim to refine the models further to develop amino acid sequences that meet various criteria for different applications. This new design tool could help address pressing societal issues in fields such as sustainability, medicine, food, health, and materials design.
Experts not involved in the study praised this research, highlighting it as a significant advance in the field. The ability to design protein sequences with desired properties opens up possibilities for various applications, including functional materials and building blocks.