The research described in this work rises from the current challenges in molecular
dynamics (MD) simulations. Although these simulations provide accurate and highresolution insights on the dynamics of biomolecular events, the timescales needed to
observe relevant events such as ligand-unbinding, protein-protein interactions and
protein folding, for instance, are not currently reachable for most scientists with
classical MD methods. Additionally, MD simulations are intrinsically complex and
high-dimensional, which makes it often difficult to elucidate and gain insights from.
To tackle the challenges in MD, an iterative protocol for ligand unbinding followed by a machine learning (ML) analysis allowed for the investigation of the
unbinding of Cyclin-Dependent Kinase 2 (CDK2) inhibitors and the long-acting
muscarinic antagonists for the human Muscarinic Receptor 3 (hMR3). This approach allowed a deeper understanding of the unbinding path and the underlying
protein-ligand interactions. This was achieved by obtaining an approximated transition state (TS) from the unbinding path and generating downhill simulations to
train two ML models to predict the outcome. ML is a powerful tool for learning to
predict from complex data. However, one of the key challenges is that many models are often considered black boxes. With explainable AI techniques it is possible
to gain insights from models and understand how the relationship between input
features and their predictions. In this work, we developed a protocol for assessing
this in a model-agnostic way and develop a framework to test this for correlated
time-series data with both 1D and 2D analytical datasets.
Additionally, a problem-tailored Hamiltonian replica exchange methodology
was also developed to aid in the research of systems mainly governed by electrostatic interactions. This is useful especially for phosphate-related enzymes where
metal ions play a role in catalysis and active site geometries. This was tested on
several systems leading to the CRISPR Cas1/Cas2 system. Results on the modelled
complex hinted at a possible two-metal ion coordination in the active site due to
major rearrangements and a K
+ ion transitioning from the bulk to form part of the
coordination.