Combining Evolution and Physics through Machine Learning to Decipher Molecular Mechanisms

QC 2024-04-23

Abstract

From E.coli to elephants, the cells of all living organisms are surrounded by a near impenetrable wall of lipids. The windows through the walls are membrane proteins - receptors, transporters and channels that confer communication, information and metabolites through the membrane. Without opening holes in the membrane, it is necessary for these proteins to alter their shapes by cycling between conformational states to transport signals or molecules. Owing to their important role as information bottle-necks, changes in their function can lead to cancer, infectious diseases, or metabolic disorders. Hence, they are important targets for drug discovery, therapeutic research and understanding the human body.

Due to the delicate thermodynamic balance of conformational states of these proteins that are modulated by external stimuli, it is difficult to trap them in experimental setups in which their native states are captured. To add to the problematic nature of their molecular mechanisms, they are too fast to kinetically trap in a certain state long enough to observe without breaking the molecular mechanism. Fast moving mechanisms makes them a good target for molecular dynamics (MD) simulations, where the movement of all atoms in the proteins is simulated over time. Although a powerful tool, modern MD simulations are not able to access long enough timescales to accurately measure macroscopic functionally relevant information, leaving a gap between simulations and reality in which many conclusions made with atomic resolution fail to translate into macroscopic phenomena, such as receptor activity, transport efficiency, mutational stability or allosteric signalling.

This work presents novel methodology that efficiently discovers and explores functionally relevant conformational states using MD simulations. By combining evolutionary information with physics using machine learning, the methodology accelerates the sampling while retaining the details of the molecular mechanism and the thermodynamic information. Additionally, the work shows how the methodology is capable of bridging the gap in resolution between experiments and simulations through the in-silico measurement of macroscopic phenomena on a microscopic scale. Moreover, it uniquely presents a framework applied to 4 studies on different target proteins of different families in which conformational change occurs, and is able to independently relate them to different types of measurements.

urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-345862