Discovery of new enzymes for degrading anthropogenic materials using knowledge graph embedding

Human-made waste is polluting water, soil and air, and is likely to increase in the coming decades. A promising avenue for its treatment is via enzymatic degradation, which has shown promise in the degradation of polyethylene terephthalate (PET), the compound that makes up plastic bottles. However, contemporary methods for detecting new enzymatic functions from microbial sources are often limited by culturing, which is restrictive to most microbes. Computational methods tackling these problems often rely on comparison to known datasets and therefore suffer from observational bias, limiting new discoveries. We approach these problems using knowledge graphs, which enable the aggregation of information from multiple sources and the application of AI algorithms to discover new links between them. As a first step, we downloaded multiple datasets of enzymes and their substrates, and applied a graph embedding algorithm to detect similarities between certain nodes (e.g., enzymes). The embedding of this massive network enables us to find alternative substrates for enzymes, to annotate the function of new enzymes, or to find known enzymes that can be applied to new substrates. Our preliminary model was tested against 1000 artificial links (non-existent in the data) for each real link. In 81% of the cases, it places the real link as its best prediction, and in 97% of cases, the real link is amongst the top 5%. Further developing this network could help us find new enzymes that are able to degrade human-made waste and potentially reduce worldwide pollution from many different sources.