Unveiling hidden intermediate states in protein folding with AI-based conditional transition clustering

Revealing the complex mechanisms of protein folding, including the transient intermediate states that govern the process, is a fundamental goal in computational biophysics. While molecular dynamics (MD) simulations generate vast amounts of data to this end, extracting a clear kinetic model from these complex, high-dimensional trajectories remains a significant challenge. We present AI-Based conditional transition clustering (CTC), a framework for analyzing MD trajectories that directly addresses the limitations of state-centric methods. Conventional approaches, such as Markov state models, rely on predefined geometric clustering or assume fixed linear dynamics, which can bias the discovery of protein conformational states. CTC operates on a "dynamics-centric" principle, defining a conformational state as a kinetically trapped region identified after analyzing the system dynamics, not before. By leveraging AI-based normalizing flows to estimate conditional transition probabilities from the MD data, CTC identifies states as "kinetic islands" with low escape probabilities. Applying CTC to protein-folding simulations successfully identifies critical intermediate and transition states, revealing folding pathways without prior assumptions about the number of states or their kinetic properties. This approach provides a more objective and physically grounded method for uncovering the complex mechanisms of biomolecular systems.