Deep Learning of Cellular Metabolic Flux Distributions Predicts Lifespan

Abstract

It is a common observation that individuals within a species age at different rates. Variation in both genetics and environmental interaction are generally thought responsible. Surprisingly, even genetically identical organisms cultured under environmentally homogeneous conditions age at different rates, implying a more fundamental cause of aging. Here we have examined the basis for lifespan variance in haploid, single-celled yeast of Saccharomyces cerevisiae. The probabilistic nature of metabolism means metabolites often, but not always, follow the same route through the metabolic network. We speculate redundancy in metabolic pathway choice is sufficient to explain lifespan variance. To interrogate the reaction flux space of S. cerevisiae we used a model of its intermediary metabolism, comprising 1,150 genes, 4,058 reactions, and 2,742 metabolites (yeast GEM_v8.5.0). We restricted traffic through the metabolic network by knocking out each of the 1,150 genes, then generated a total of 406,500 flux distributions spanning the solution space of the resulting 812 viable mutants. We used replicative life span (RLS) data for the 812 viable mutants, corresponding to 66,400 individual cells. Four approaches were then employed to test whether reaction flux configuration could be used to predict lifespan: Principal Component Analysis (PCA) in conjunction with non-linear modeling of RLS; deep learning of RLS using either a Regression Neural Network (RNN) or a Classification Neural Network (CfNN); and deep learning using a convolutional neural network (CNN) following conversion of flux distributions to pixelated images. The four approaches reveal a core network of highly correlated reactions controlling aging rate that is sufficient to explain all lifespan variance. It includes biosynthetic pathways encompassing ceramides, monolysocardiolipins, phosphoinositides, porphyrin and glycerolipids. Our data lead to two novel conclusions. First, variance in the replicative lifespan of S. cerevisiae is an emergent property of its metabolic network. Second, there is convergence among metabolic configurations toward three meta-stable flux states - one associated with extended life, another with shortened life, and a third with wild type life span. One Sentence Summary: Traffic routes and rates through the metabolic network of S. cerevisiae fully account for variance in replicative lifespan.