Â鶹ÒùÔº

In a study published in , researchers at the University of Wisconsin–Madison introduced a deep learning method capable of automatically identifying transition states in protein conformational changes, a key process that underpins many biological functions.

This new tool promises to accelerate the study of biomolecular dynamics and could have wide-reaching applications in drug design, biomolecular engineering, and materials science.

This study is a collaborative effort between Prof. Xuhui Huang's group (Department of Chemistry) and Prof. Sharon Li's group (Department of Computer Sciences) at the University of Wisconsin–Madison.

Transition state identification has long been considered the "holy grail" in chemistry. Unlike chemical reactions, biomolecular conformational changes, such as protein folding or binding to other molecules, involve multiple metastable intermediate states, giving rise to numerous transition states situated at the free energy barriers within a complex landscape.

Despite decades of research, existing methods have only been able to locate transition states between pairs of metastable states. The simultaneous and automatic identification of all transition states in biomolecular processes has remained a major challenge.

The new technique, named TS-DAR (Transition State identification via Dispersion and vAriational principle Regularized neural networks), overcomes these challenges by leveraging a deep learning framework inspired by out-of-distribution (OOD) detection—a concept from artificial intelligence (AI) used to identify data that deviates from typical patterns.

The key breakthrough of TS-DAR is its ability to treat transition states as OOD data—rare structures located at the free energy barriers between metastable conformations. The method works by embedding molecular dynamics (MD) data into a hyperspherical latent space, where it can efficiently detect and isolate these sparsely populated transition states.

This approach provides a comprehensive, end-to-end pipeline for studying protein dynamics and identifying all transition states involved in biomolecular processes.

"Identifying transition states is one of the most challenging and important tasks in studying protein dynamics," said Prof. Xuhui Huang. "TS-DAR is the first method capable of automatically capturing all transition states at once from MD data, enabling a much deeper understanding of the underlying molecular processes."

The research team tested TS-DAR on a range of systems, including the translocation of a DNA motor protein (AlkD) along DNA. In each case, TS-DAR outperformed traditional methods in both accuracy and efficiency.

Notably, in the AlkD system, the method revealed new insights into the role of protein-DNA hydrogen bonds, which play a critical role in determining the rate-limiting step of AlkD's translocation—an important process in DNA repair.

With its ability to detect transition states in complex biomolecular systems, TS-DAR represents a significant advancement in the study of molecular dynamics.

The framework's potential to accurately model highly dynamic processes could also pave the way for the development of generative AI models, offering new avenues for predicting and manipulating biomolecular dynamics.