AI-Enhanced Molecular Dynamics Simulations for Protein Folding and Drug Binding Prediction

Although the common practice in drug discovery is to screen compounds against a rigid structure for faster evaluations of multiple candidates, the robust prediction of drug binding should include the flexibility of the protein targets. For pharmaceuticals that induce the desired state of a protein, the receptor-drug dawning process can often be captured within protein conformational fluctuations about a bound crystal structure. Molecular Dynamics (MD) simulation has become an effective approach to see such fluctuations and has matured to provide a complement to and a probe for X-ray crystallography.

One-tenth nanosecond MD simulations produce sufficient flexibility information about the structure and configuration of drug binding which are used for ensemble docking against a cluster of the sampled structures. Docking methods have improved to properly account for topology and conformational changes of a protein through models such as the rigid receptor, flexible compound, induced fit, and ensemble docking. The first three methods require extensive sampling and are computationally more expensive than rigid receptor docking. Meanwhile, parallelization is possible at either conformation or compound-level with respect to the respective dimension of the problem but cannot be performed at both levels since the dimensions of the problem are critically diverse in MDO.

There has been considerable interest in deriving and implementing accurate, fast, and flexible soft potentials in machine learning (ML) frameworks. However, these approaches have either been serial implementations with no speedup with respect to numerical relaxation or did not fully utilize end-to-end optimization through gradients. Since a network trained on a protein family would not be applicable to one outside the family,. Molecular MD would be a better training starting point as it ensures better physical prior during learning until reaching a steady state.