Shulou(Shulou.com)11/24 Report--

CTOnews.com, April 24 (Xinhua) protein is the foundation of life and the inspiration for many new materials. However, the known proteins in nature are only the tip of the iceberg, and there are countless undiscovered proteins waiting to be explored. How to design novel proteins with specific structures and functions quickly and effectively is a great challenge.

To solve this problem, researchers at the Massachusetts Institute of Technology (MIT) have used artificial intelligence to develop a way to produce new proteins that transcend nature. They created an algorithm based on machine learning that can predict the amino acid sequence that can form the corresponding protein according to the preset structural goal. These sequences are not only similar to known proteins, but also innovative and unique. These proteins can be used to make materials with specific mechanical properties (such as stiffness or elasticity) to replace oil or ceramics as raw materials, greatly reducing the carbon footprint.

The senior author of the study, which will be published in the latest issue of Chemistry, is Markus Buehler, a professor at the MIT School of Engineering, a professor in the departments of Civil and Environmental Engineering and Mechanical Engineering, and a member of the MIT-IBM Watson artificial Intelligence Laboratory. He said that this method can provide new solutions for a variety of applications, such as biomedicine, material science, food preservation and so on. "in terms of proteins that transcend nature, this is a huge design space that cannot be solved artificially," he said. "We need to understand the language of life, how to encode amino acids in DNA, and how they can be combined into protein structures. This is impossible before deep learning."

The study was co-authored by Bo Ni, a postdoctoral fellow at Buller Lab, and David Kaplan, a professor in the department of bioengineering at Tufts University University and dean of the school of engineering.

Using the new machine learning model, proteins are long chains of amino acids that fold into a three-dimensional shape. The sequence of amino acids determines the structural characteristics of proteins, which in turn affect the mechanical properties of proteins. Although scientists have discovered thousands of proteins formed by natural selection, they estimate that many amino acid sequences remain undiscovered.

In order to speed up the process of protein discovery, scientists have recently designed some deep learning models that can predict the three-dimensional structure of proteins according to a given amino acid sequence. However, in turn, it is a more complex problem to predict the corresponding amino acid sequence according to the given structural target.

Buller and his colleagues were able to deal with this problem because they took advantage of a new machine learning model called the attention-based diffusion model (CTOnews.com Note: attention-based diffusion model). Buller explains that attention-based models are important for protein design because they can learn and capture long-distance relationships. This is common in proteins because a mutation in a long amino acid sequence can have a big impact on the overall design.

The learning process of diffusion model is to restore the original data by adding "noise" to the training data, and then removing "noise". These models can generate high-quality and realistic data and can be adjusted according to specific design goals. Therefore, they are more suitable to meet the design requirements than other models.

Using this architecture, researchers have developed two machine learning models that can predict amino acid sequences that meet specific structural objectives. In the field of biomedicine, having a completely unknown protein can be problematic because its properties are unclear, Buller said. However, in some applications, it may be necessary to design a new type of protein with similar properties but different functions to those existing in nature. By using the model they developed, a series of proteins can be generated and their design can be controlled by adjusting some parameters to achieve customized requirements.

Different amino acid folding patterns in proteins, called secondary structures, will lead to different mechanical properties. For example, proteins with a-helix structure tend to be elastic, while proteins with a β-fold structure are usually rigid. Combining both α-helix and β-folding structures in a protein can create materials that are both elastic and strong, just like silk.

The researchers created two models, one working at the overall structural level and the other at the amino acid level. Both models can combine amino acids to produce proteins. In the first model, users only need to enter the percentage of different structures they want, such as 40% α-helix and 60% β-folding, and the model will generate sequences that meet these requirements. In the second model, the user has to specify not only the percentage, but also the order of the amino acid structure, thus having more control over the final product.

To verify that the resulting protein meets the expected specifications, the researchers linked the developed model to an algorithm that can predict protein folding. They used this algorithm to determine the three-dimensional structure of the generated protein, then calculated the corresponding mechanical properties, and compared it with the preset design requirements. This allows them to verify that the designed protein meets the desired specifications.

Innovative and reliable designs to assess the effectiveness of their models, the researchers compared newly generated proteins with known proteins with similar structural properties. They found that many of the resulting proteins had about 50% to 60% overlap with known amino acid sequences, indicating that they were synthesizable. In addition, the models produce completely new sequences, demonstrating their ability to design new proteins.

Buller says the similarity between the generated and known proteins suggests that the designed proteins are likely to be realistic and synthesizable. In order to verify the reliability of the design protein, the researchers tried to deceive the model with some physically impossible design goals. However, the model does not produce unlikely proteins, but the ones that are closest to a viable solution. This result shows that the model is robust, and even if unrealistic design specifications are given, the closest feasible solution can be found.

Ni Bo pointed out that machine learning algorithms can discover hidden relationships in nature. This ability gives researchers confidence that the resulting protein is likely to be realistic and synthesizable.

In the next step, the researchers plan to verify some of the newly designed proteins by synthesizing them in the laboratory. In addition, they plan to further improve and refine their models so that they can design amino acid sequences that meet more conditions, such as specific biological functions.

The ultimate goal is to develop a multi-functional platform that can generate a variety of protein designs for a variety of applications, including biomedicine and materials science. Buller stressed that these applications need to go beyond the solutions offered by nature, such as sustainability, medicine, food, health and material design. Therefore, newly developed design tools can play an important role in solving these problems.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.