Net2Onto: Leveraging molecular networks and ontology structure for predicting novel genes that play a role in understudied functions, traits, and diseases

Despite numerous experiments, only about half of the ~20k protein-coding genes have any experimental evidence linking them to at least one biological function and only a quarter to any phenotype and disease. Rapidly discovering these links for un(der-)characterized genes is critical for closing enormous gaps in our current knowledge about the molecular basis of physiological processes, disease etiology, and clinical outcomes. To address this need, we and others have developed machine learning (ML) methods that can capture interaction patterns in molecular networks to predict novel genes associated with a variety of function, trait, and disease contexts. However, the reliance of these methods on previously-annotated genes severely limits their ability to make predictions for contexts with no or very few already known annotations. Methods such as clusDCA and DeepGOZero overcome this limitation by using the relationships between functions captured in the Gene Ontology to leverage information about well-studied functions to improve gene prediction performance for under-studied functions and, further, allow for predicting genes for functions with no known gene annotations. Here, we present Net2Onto, a general-purpose neural network framework that extends these methods to make novel gene discoveries about un(der-)studied contexts by learning a mapping from representations of genes in a molecular interaction network to an embedding space with terms in a biomedical ontology. Our framework combines two desirable properties of previous methods: i) like clusDCA, it uses a representation of the ontology graph structure, ii) like DeepGOZero, it uses a neural network to learn a mapping between genes and terms, which allows for the model to capture nonlinear relationships. Net2Onto extends the application of this approach to the Mondo disease ontology and the Phenio phenotype ontology in addition to the Gene Ontology biological processes. Net2Onto achieves performance comparable to baseline network-based ML models, and through rigorous evaluations where all annotations to certain terms are completely held-out, we show that it has the capability to make zero-shot predictions for terms that were not seen during training. Using this framework, we compared various ways to embed the network and the ontology, and explored how term properties correlated with model performance. Overall, Net2Onto is a powerful and general framework for leveraging networks and ontologies to discover novel genes related to functions, phenotypes, and diseases.