Moving from network representations to hierarchical models of cell structure and function

We have shown very promising results in the hierarchical analysis of physical and genetic interaction networks—i.e., that networks harbor rich structure which is modular, hierarchical and multi-scale. In particular, we were able to recover ~60% of the cell component hierarchy recorded by the Gene Ontology (GO) from literature, but in a de novo fashion directly from physical and genetic network [Dutkowski et al. Nature Biotechnology 2013]. The resulting Network-eXtracted Ontology, which we called NeXO, provides a structured hierarchical interpretation of network data which will in most cases be vastly preferable to flat lists of interaction (a.k.a. interaction ‘hairballs’) or flat lists of network clusters/complexes. 

Figure 1. Automated assembly and alignment of gene ontologies.

Based on this progress, in the present review period we have begun detailed construction of hierarchical network models for different aspects of cell biology important to cancer and aging. The first such model to be published is of autophagy, a collaboration with the laboratory of Dr. Suresh Subramani at UCSD [Kramer et al. Molecular Cell 2017]. The resulting hierarchy contains 220 distinct autophagy subsystems, approximately half of which were previously unknown. We have thus far confirmed roles for Gyp1 at the phagophore-assembly site, Atg24 in cargo engulfment, Atg26 in cytoplasm-to-vacuole targeting, and Ssd1, Did4, and others in selective and non-selective autophagy. The autophagy hierarchy is available at We presently working on a hierarchical model of DNA repair built from protein and genetic interaction data using related methods to those above.

Active Interaction Mapping reveals the hierarchical organization of autophagy. Graphical Abstract. 

Given these hierarchical cell structures we were learning to construct, a natural next step was to match hierarchical structure with hierarchical function. Could we use the hierarchical cell structure to model the hierarchical effects of genetic variants and perturbagens, not only on individual proteins, but on protein complexes, pathways and organelles at multiple cellular scales? We have thus far published two attempts to answer this question [Yu et al. Cell Systems 2016; Ma et al. Nature Methods 2018].

In this latest work by Ma et al., we developed a whole-cell structure/function model called DCell, a deep neural network encoding both the structure and phenotypic growth behavior of a budding yeast cell. In this model, a deep neural network is constructed whose structure is guided by prior knowledge of the cellular hierarchy of processes and components. Each subsystem in the hierarchy is represented by a group of neurons that are restricted in their connectivity, such that these neurons only take input from the neurons of child subsystems and only provide output to the neurons of parent subsystems. This configuration thus treats the cell as a type of information processing system captured mathematically by a neural network. DCell was trained on a large compendium of ~12 million genotype-phenotype measurements in budding yeast, in which each genotype comprised a small combination of deleted genes (typically two). We found that DCell achieves significant gains in accuracy in phenotype prediction relative to earlier approaches. Moreover, unlike standard “black box” neural networks, the simulations performed by DCell are tied to an extensive hierarchy of internal biological subsystems with states that can be queried. This “visible” aspect enables in silico investigations of the molecular mechanisms underlying each genotype-phenotype prediction.

Figure 1. Modeling system structure and function with visible learning. From: Using deep learning to model the hierarchical structure and function of a cell. [Ma et al. Nature Methods 2018]

Cover articels: Yu et al. Cell Systems 2016Ma et al. Nature Methods 2018