Scientists have discovered a new approach to forecasting how proteins alter their shapes during their activities, which is crucial for grasping their roles in living organisms. Although recent advancements in artificial intelligence (AI) have facilitated the prediction of protein structures in their inactive state, understanding how these proteins move remains problematic due to a lack of sufficient experimental data on protein dynamics to effectively train neural networks.
Scientists have discovered a new approach to forecasting how proteins alter their shapes during their activities, which is crucial for grasping their roles in living organisms. Although recent advancements in artificial intelligence (AI) have facilitated the prediction of protein structures in their inactive state, understanding how these proteins move remains problematic due to a lack of sufficient experimental data on protein dynamics to effectively train neural networks.
In a recent study published in the Proceedings of the National Academy of Sciences on August 20, Rice University’s Peter Wolynes, along with peers in China, combined insights about protein energy landscapes with deep-learning techniques to predict these movements.
Their technique enhances AlphaFold2 (AF2), a tool that predicts static protein structures, by training it to concentrate on “energetic frustration.” Proteins have evolved to minimize energy conflicts among their components to reach their stable configurations. Persistent conflicts indicate regions of frustration.
“Starting from the predicted static ground-state structures, the new method generates alternative structures and pathways for protein motions by initially identifying and then progressively amplifying the energetic frustration features in the input multiple sequence alignment sequences that encapsulate the protein’s evolutionary history,” explained Wolynes, the D.R. Bullard-Welch Foundation Professor of Science and co-author of the study.
The researchers applied their method to the protein adenylate kinase and found that their predicted movements aligned with experimental observations. They also successfully forecasted the dynamic changes of other proteins that undergo significant shape alterations.
“Predicting the three-dimensional structures and movements of proteins is essential for understanding their functions and for developing new medications,” Wolynes noted.
The study also evaluated how AF2 operates, demonstrating that merging physical insights regarding the energy landscape with AI not only aids in predicting protein movements but also clarifies why AI tends to overestimate structural stability, resulting in predictions that lean towards the most stable configurations.
The energy landscape theory, which Wolynes and his colleagues have explored for decades, is a crucial element of this technique. However, recent AI algorithms have been conditioned to predict solely the most stable protein forms, overlooking the various shapes proteins can adopt during their functions.
The theory posits that while evolution has shaped the protein’s energy landscape to facilitate folding into optimal structures, deviations from a perfectly guided landscape – termed local frustration – play a vital role in functional protein dynamics.
By identifying these frustrated areas, the researchers trained the AI to disregard them when making predictions, enabling it to more accurately forecast alternative protein configurations and their functional movements.
Utilizing a frustration analysis tool within the energy landscape framework, the team pinpointed flexible, frustrated regions within proteins.
By adjusting the evolutionary insights from aligned protein family sequences employed by AlphaFold and aligning them with the frustration scores, they trained the AI to identify these frustrated zones, leading to precise predictions of alternative configurations and transitions between them, said Wolynes.
“This research highlights the importance of not neglecting or eliminating physics-based methods in the post-AlphaFold age, where the focus has been on unbiased learning from experimental data without theoretical guidance,” Wolynes emphasized. “Fusing AI with biophysical knowledge will greatly influence future applications, including drug formulation, enzyme development, and understanding mechanisms behind diseases.”
Other contributors to the study include Xingyue Guana, Wei Wanga, and Wenfei Lia from the Department of Physics at Nanjing University; Qian-Yuan Tang from the Department of Physics at Hong Kong Baptist University; Weitong Ren from the Wenzhou Key Laboratory of Biophysics at the University of Chinese Academy of Sciences; and Mingchen Chen from the Changping Laboratory in Beijing.
The research received support from the National Natural Science Foundation of China, the Wenzhou Institute at the University of Chinese Academy of Sciences, the Hong Kong Research Grant Council, and the U.S. National Science Foundation-funded Center for Theoretical Biological Physics at Rice University.
