Applying machine learning to investigate lipid domain formation behaviors in the presence of amyloidogenic proteins.

Abstract

Listen

Translate article

Cell membranes are not homogeneous but contain lipid nanodomains distributed on both leaflets of the lipid bilayers. Using microsecond molecular dynamics (MD) simulations, we have simulated phase-separated lipid bilayers (raft membranes) containing lipid nanodomains that mimic the heterogeneous structures of neuronal plasma membranes. Our recent studies indicate that these lipid nanodomains, e.g., the mixed liquid-ordered (Lo) and liquid-disordered (Ld) boundaries or Lod and ganglioside-cluster, represent non-specific and specific membrane damage targets of amyloidogenic protein, e.g., tau, aggregates (oligomers), in raft membranes. However, a robust computational tool to classify lipids into these lipid nanodomains and quantitate the spatial and temporal evolutions of the domain structures in the presence of amyloidogenic oligomers is not available. In this study, we have modified two unsupervised Machine Learning (ML) algorithms, Non-negative Matrix Factorization (NMFk) and K-Mean Clustering, to investigate lipid domain formation behaviors. These algorithms are currently used in mutation analysis of cancer genomes and feature extractions. In this work, NMFk was used to decompose our time-resolved lipid-lipid proximity data matrix into two latent “signature” and “activity” matrices, which contain a set of probability configurations and the time-dependent activity of those configurations, respectively, of the lipids participating in each nanodomain, e.g., Lo and Lod. On the other hand, by implementing additional spatial information to the dataset used in NMFk, the K-Mean Clustering sorts lipids into nanodomains and determines the latent temporal and spatial evolutions of each domain. Hence, by combining both ML algorithms, this study provides new insights into understanding protein-induced membrane damage mechanisms of amyloidogenic oligomers associated with lipid domain disruption behaviors that may contribute to the pathogenesis of Alzheimer's. (Supported by NSF [OAC 153159] and Welch Foundation [W-2057-2021-327].)