| Details: |
Molecular dynamics (MD) simulations are used extensively to study the mechanisms of biological processes in atomistic resolution. But most physiological events, e.g. drug-target binding and protein folding, take place at beyond millisecond timescales, while current computing facilities allow simulations only up to a few microseconds. A number of enhanced sampling algorithms e.g umbrella sampling, metadynamics etc. have been developed in the past few decades to accelerate conformational sampling by applying external biasing potential. The accuracy and efficiency of these algorithms are sensitive to the choice of collective variable (CV), a low dimensional space along which the bias is applied. Intuitive CVs, e.g. distances, angles, etc. are often insufficient to adequately sample the conformational landscape. Machine learning algorithms can play a key role in addressing these challenges. We demonstrated that collective variables constructed using deep neural networks with a generic and system-agnostic feature space provides an accurate free energy surface for complex molecular systems e.g. protein folding and ligand binding. Using it in combination with the novel On-the-fly probability enhanced sampling (OPES) flooding algorithm, the kinetic properties can also be recovered. Furthermore, a multi-task machine learning approach that combines supervised and unsupervised learning can identify the minimum free energy pathway in the rugged conformational landscape of complex molecular systems. We demonstrated a successful application of this technique in calculating a detailed free energy profile for ligand binding to a G-protein coupled receptor. Through these examples I will illustrate the significant potential of deep learning algorithms in studying biomolecules relevant for drug discovery. |