Confronting pitfalls of AI-augmented molecular dynamics using statistical physics

Abstract

Artificial intelligence (AI) based approaches have had indubitable impact across the sciences through the ability to make sense of data. Recently AI has also seen use for enhancing the efficiency of molecular simulations, wherein AI derived slow modes are used to accelerate the simulation in targeted ways. However, while typical fields where AI is used are characterized by a plethora of data, molecular simulations per construction suffer from limited sampling and thus limited data. As such the use of AI in molecular simulations can suffer from a dangerous situation where the AI optimization could get stuck in spurious regimes, leading to incorrect characterization of the reaction coordinate for a given problem at hand. When such an incorrectly characterized reaction coordinate is then used to perform additional simulations or even experiments, one could start to deviate further and further from the ground truth. To deal with this problem of spurious AI solutions, here we report a new and automated algorithm using ideas from statistical mechanics. It is based on the notion that a more reliable AI solution for many problems in chemistry and biophysics will be one that maximizes the time scale separation between slow and fast processes. To learn this timescale separation even from limited data, we use a maximum path entropy or Caliber based framework. We show the applicability of this automatic protocol for 3 classic benchmark problems. Here we capture the conformational dynamics of a model peptide, ligand unbinding dynamics from a protein and the extensive sampling of the folding/unfolding energy landscape of a GB1 peptide. We believe our work will lead to increased and robust use of trustworthy AI in molecular simulations of complex systems.

Artificial intelligence techniques for integrative structural biology of intrinsically disordered proteins

“…Current AI/ML applications (barring a few like [44,39]) tend to use fitting procedures in a blind manner, without much physical bearing, or paying attention to the underlying statistical physics of the system of interest. The resulting fitting procedure can end up overfitting and may not generalize to fully leverage the power of ML/AI in other domains [43,21]. In particular, transferring a ML/AI model learned across simulations can be challenging.…”