Researchers teach artificial intelligence about protein folding frustration

Scientists have discovered a new way to predict how proteins change shape when they function, which is important for understanding how they work in living systems. While recent artificial intelligence (AI) technologies have made it possible to predict what proteins look like at rest, understanding how they move remains difficult because there is not enough direct data from experiments on protein movements to train neural networks.

In a new study published in the Proceedings of the National Academy of SciencesPeter Wolynes of Rice University and his colleagues in China combined information about protein energy landscapes with deep learning techniques to predict these movements.

Their method improves AlphaFold2 (AF2), a tool that predicts static protein structures, by teaching it to focus on “energy frustration.” Proteins have evolved to minimize energy conflicts between their parts, so that they can be channeled toward their static structure. When conflicts persist, there is said to be frustration.

“From predicted static ground-state structures, the new method generates alternative structures and pathways for protein motions by first finding and then progressively improving energy frustration features in input multiple sequence alignment sequences that encode the protein’s evolutionary development,” said Wolynes, the D.R. Bullard-Welch Foundation Professor of Science and co-author of the study.

The researchers tested their method on the protein adenylate kinase and found that its predicted motions matched the experimental data. They were also able to predict the functional motions of other proteins that change shape significantly.

“Predicting the three-dimensional structures and motions of proteins is essential for understanding their functions and designing new drugs,” Wolynes said.

The study also examined how AF2 works, showing that combining physical knowledge of the energy landscape with AI not only predicts how proteins move, but also explains why AI overestimates structural integrity, leading to only the most stable structures.

Energy landscape theory, which Wolynes and his collaborators have worked on for decades, is a key part of this method, but recent AI codes have been trained to predict only the most stable protein structures and ignore the different shapes proteins can take when they function.

Energy landscape theory suggests that while evolution has sculpted the energy landscape of proteins where they can fold into their optimal structures, deviations from a perfectly channeled landscape that otherwise guides folding, called local frustration, are essential to the functional movements of proteins.

By identifying these frustrated regions, the researchers taught the AI ​​to ignore these regions to guide its predictions, allowing the code to accurately predict alternative protein structures and functional motions.

Using a frustration analysis tool developed within the energy landscape framework, the researchers identified frustrated and therefore flexible regions in proteins.

Then, by manipulating the evolutionary information in the aligned protein family sequences used by AlphaFold and according to the frustration scores, the researchers taught the AI ​​to recognize these frustrated regions, enabling accurate predictions of alternative structures and pathways between them, Wolynes said.

“This research highlights the importance of not forgetting or abandoning physics-based methods in the post-AlphaFold era, where the focus has been on agnostic learning from experimental data without any theoretical input,” Wolynes said. “Integrating AI with biophysical knowledge will have a significant impact on future practical applications, including drug design, enzyme engineering, and understanding disease mechanisms.”

Other authors include Xingyue Guana, Wei Wanga and Wenfei Lia of the Department of Physics at Nanjing University; Qian-Yuan Tang of the Department of Physics at Hong Kong Baptist University; Weitong Ren of the Wenzhou Key Laboratory of Biophysics of the University of Chinese Academy of Sciences; and Mingchen Chen of the Changping Laboratory in Beijing.