Protein-lipid interactions play a key role in regulating the activity of membrane-binding and integral membrane proteins. However, the specific interactions between proteins and lipid molecules remain challenging to characterize in molecular detail, especially at large scale. We show how high-throughput coarse-grained molecular dynamics simulations and machine learning can fill this gap, providing realistic models at both the individual protein and family-wide level.
Extensive molecular dynamics simulations and complementary experimental approaches are used to characterize membrane binding by the catalytic core domains of phospholipase C γ1, revealing previously unknown sites of lipid interaction on this clinically important phospholipase. Direct evidence is provided that PLCγ1 is autoinhibited through obstruction of its membrane-binding surfaces. Knowledge of the critical sites of membrane interaction extends the mechanistic framework for activation, dysregulation, and therapeutic intervention.
For the pleckstrin homology (PH) domain family of membrane binding domains, the systematic simulation of the interactions of 100 mammalian PH domains with PIP-containing membranes is reported. The observed PIP interaction hotspots recapitulate crystallographic binding sites, revealing a number of insights; including the patterns of PIP interaction across the family, details of secondary binding sites, clustering of anionic lipids and membrane-bound orientations. This provides a global view of PH domain/membrane association involving multivalent association with anionic lipids.
Going beyond MD simulations, we demonstrate that graph neural networks trained on coarse-grained MD simulation data can predict phosphoinositide lipid interaction sites on PH domain structures.
The predictions are comparable to the results of simulations and require only seconds to compute. Comparison with experimental data shows that the model can predict known phosphoinositide interaction sites and can be used to form hypotheses for PH domains for which there is no experimental data. This opens up the possibility to put powerful machine learning based tools for predicting protein-lipid interactions in the hands of membrane biologists. Ongoing work to extend these predictions to other families of membrane proteins is discussed.
In a parallel strand of research, the use of electron capture for charge reduction (ECCR) in native mass spectrometry on Orbitrap platforms has been explored. Native ECCR was performed on the bacterial chaperonin GroEL and megadalton scale adeno-associated virus (AAV) capsid assemblies on the Q Exactive UHMR mass spectrometer. Charge reduction of AAV8 capsids by up to 90% pushes signals well above 100,000 m/z and enables charge state resolution and mean mass determination of these highly heterogeneous samples, even for capsids loaded with genetic cargo. With minor instrument modifications the UHMR can detect charge reduced ion signals beyond 200,000 m/z. This work demonstrates the utility of ECCR for deconvolving heterogeneous signals in native mass spectrometry and presents the highest m/z signals ever recorded on an Orbitrap instrument, opening up the use of Orbitrap native mass spectrometry for heavier analytes than ever before.
