On the trail of proteins on a roller coaster: from mapping conformational free-energy surfaces to computational drug design

Abstract

Molecular structures are not static; they explore vast conformational space. One of the most successful models in computational chemistry, the potential-energy surface and free-energy landscape, lets researchers study conformational dynamics on maps that resemble a tourist’s guide. This short note traces four decades of my roller-coaster ride across this molecular landscape, from small-ring conformational analysis to protein dynamics and drug discovery. A foundational study mapped the full network of interconversion pathways of the C4–C12 cycloalkanes and led to Low Mode Search, based on using a few low-frequency normal modes that capture biologically relevant protein motions. On the “proteographical” map, contours represent the probability of different conformations, expressed in energy units as the free energy. Minimum-free-energy valleys correspond to distinct, biologically relevant conformations, and proteins perform the dance of life by moving from valley to valley choreographed by cell’s biochemical pathways. Many diseases arise from altered terrain that disrupts this choreography, and drug discovery aims to restore the original dance. Molecular dynamics simulations are essential for understanding free-energy landscapes, but conventional MD is too slow on commercial hardware for practical drug discovery. Biased MD addresses this limitation by accelerating dynamics in a small subspace defined by a few biologically relevant collective variables. In this CV space, hundreds of microseconds of hypothetical unbiased simulations can be replaced by hundreds of nanoseconds of biased simulations, yielding a three-order-of-magnitude speedup on GPU desktop computers. We applied one-dimensional path-CV simulations in the meta-eABF framework to study conformational transitions in several systems, including the oncogenic JAK2-V617F pseudokinase domain. A different approach, our co-vibrational modes theory, extracts dynamic information from two static X-ray structures through a weighted superposition of low-frequency normal modes. Using covib modes, we reproduced the seesaw helix motion that dominates the conformational transition between wild-type and mutant JAK2-V617F structures, which no single mode captured. Covib modes can also “shake” a single protein structure to expose cryptic pockets that are closed in X-ray structures but may open during natural protein motion and serve as drug targets. Finally, although machine learning methods such as AlphaFold⁷ have transformed protein structure prediction, AI-based simulation of protein motion remains in its infancy. Our reinforced molecular dynamics (rMD) method was designed to give the otherwise physics-agnostic latent space of neural networks real physical meaning. We showed that a network can be trained by optimizing two loss functions simultaneously, so the latent space can be replaced by the free-energy map obtained from meta-eABF simulation. We demonstrated rMD on the open-to-closed transition of cereblon (CRBN), central to molecular-glue degraders. rMD acts as an afterburner for a completed simulation: one can draw an arbitrary path on the free-energy map and generate the corresponding conformational transition in seconds. Together, these methods bring large-scale exploration of biologically relevant protein motions into computational drug discovery on affordable GPU-enabled desktop computers.