Research Projects

Biological membranes, or lipid bilayers, are self-assembled compartments of myriad phospholipids. Although planar in appearance at short lengths, their curvature response underscores key cellular functions, while their distortion can trigger a range of diseases. It is recognized now that bilayer curvature responses are driven by membrane associated proteins. However, a priori prediction of membrane shape architecture to protein sequence remains an unmet challenge. Inspired by our recent proof-of-concept work, this proposal seeks to leverage the structural and thermodynamic information derived from atomistic and coarse-grained molecular simulations to develop an artificial intelligence (AI) platform for membrane curvature prediction directly from the associated protein. Molecular simulations of three major classes of membrane proteins in their closest phospholipid bilayer milieu will be used to extract necessary training features for the model; these will putatively include the protein’s sequence and charge, components of interaction strength, protein configurational entropy, and the entropy of bilayer mixing at equilibrium. The features will underlie the development of a generalized deep neural network for direct prediction of membrane curvature geometry from the protein sequence. The method will deploy multi-layer graph neural networks (GNNs) that incorporate message passing algorithms for integrating node information and topological structure. Overfitting will be carefully monitored with algorithms such as propagation- and noisy- regularization. The resultant module is expected to significantly aid mechanistic studies of protein driven membrane shape responses. Feedback received from laboratory users will be used to further refine the training. The refined AI platform will be of potential commercial value in terms of licensing to computational software and drug discovery industries.