Relational Graph Convolutional Networks for Phenotypes

Relational-GCN for phenotypes

• Category: Machine Learning

• Lead: Claus Weiland (0000-0003-0351-6523)

• Members:

• One-line description: Goal of topics Various types of digitized biological data like sequences, annotations, occurrences and traits can be modelled as comprehensive knowledge graphs (involving an ecosystem of identifiers, registries, protocols, curation ontologies like BCO..), capturing interactions and similarity between taxa, and provide in this way a use case for Machine Learning-based knowledge discovery. During Biohackathon 2019, we followed an approach employing Relational Graph Convolution Networks (R-GCNs; Schlichtkrull et al.; arXiv:1703.06103) to investigate semantic relations in a large biodiversity knowledge graph (FLOPOknb; semantics.senckenberg.de/sparql)

• Goal: Employ R-GCNs in the context of Knowledge Graphs, specifically for link prediction and entity classification. Likewise we want to evalute R-GCNs as standalone models for downstream analysis (compared e.g. to WORD2Vec-like embedding-based approaches).

• Link to the project page (if any): https://github.com/cp-weiland/biohackathon2019

• Slack channel: none

• STATUS: A R-GCN model was implemented, which refers to the KG as a Multigraph with entities as nodes and relations as labelled edges. The node classification task was finalized, we are currently working on the link prediction task (which requires an additional decoding step for the prediction of new links/insertion of new triples).