Machine Learning in Transition State Searching for Complex Biomolecules

Transition state is the key concept for chemists to understand and fine-tune the conformational changes of large biomolecules. Due to its short residence time, it is difficult to capture a transition state via experimental techniques. Characterizing transition states for a conformational change therefore is only achievable via physics-driven molecular dynamics simulations. However, unlike chemical reactions which involve only a small number of atoms, conformational changes of biomolecules depend on numerous atoms and therefore their coordinates in our 3D space. The searching of transition states for them will inevitably encounter the curse of dimensionality, i.e. the reaction coordinate problem, which invoked the invention of various algorithms for solution. Recent years have witnessed a rapid emergence of new machine learning techniques and the incorporation of some of them into the transition state searching methods. Here, we first review the design principle of representative transition state searching algorithms, including the collective-variable(CV)-dependent Gentlest Ascent Dynamics (GAD), Finite Temperature String, Fast Tomographic, Travelling-salesman based Automated Path Searching (TAPS) and the CV-independent Transition Path Sampling (TPS). Then, we focus on the new version of TPS that incorporates reinforcement learning for efficient sampling and clarify the suitable situation for its application. Finally, we propose a new paradigm for transition state searching-invent new dimensionality reduction techniques that preserve transition state information and combine them with GAD.