PyKEEN is a framework, which integrates several approaches to compute knowledge graph embeddings (KGEs). We demonstrate the usage of PyKEEN in an biomedical use case, i.e. we trained and evaluated several KGE models on a biological knowledge graph containing genes' annotations to pathways and pathway hierarchies from well-known databases. We used the best performing model to predict new links and present an evaluation in collaboration with a domain expert?.
Knowledge graphs (KGs) have been adopted by various research elds (e.g.,
Semantic Web, bioinformatics) to represent factual information. Examples of
KGs are DBpedia [
Because the software ecosystem for KGEs remains limited, we have developed
the KEEN Universe [
In this demonstration paper, we present a link prediction use case from the
biomedical domain that accompanies our resource paper at the ISWC 2019
Conference [
PyKEEN provides the functionalities to train and to evaluate KGEs, and it provides an inference work ow that assists users to predict novel links. PyKEEN consists of two layers: the con guration layer and the learning layer. Con guration Layer The con guration layer assists users in specifying their KGE experiments' datasets, models, hyper-parameters, training procedures, and evaluation procedures. Experiments can either be de ned programmatically or by using the interactive command line interface (CLI) via a terminal. The CLI ensures that experiments are con gured correctly, and in case users provide an invalid input, the CLI informs the users and provides an example of a correct input.
Learning Layer The learning layer trains a model in training mode based on a user de ned set of hyper-parameters or nds suitable hyper-parameter values in hyper-parameter optimization (HPO) mode.
Inference Work ow The inference work ow generates for a user de ned set of entities and relations all possible triple permutations (users can specify to exclude re exive triples of the form (e,r,e)). The inference work ow exports a le containing the triples and their predicted scores where the predictions are sorted according to their scores. 3
Biological pathway databases have been generated and used in the classical analysis of -omics data, but their formulations as KGs are not amenable to classical machine learning approaches for classi cation, clustering, or predictive modeling. Here, we trained and evaluated three KGE models (i.e., TransE, TransR and ComplEx) before selecting the best performer for prediction of novel roles of genes in pathways and evaluation by a domain expert.
Training was conducted using four datasets: three that comprise links
between genes and pathways from disparate pathway databases (i.e., KEGG [
While ComplEx performed well, the poor performance of TransE may be due its poor abilities to handle the high cardinality (N-M) relations in the data. Due to time constraints, only 5 iterations of HPO were performed for TransR, so its results may also be improved.
Results Due to the large number of predictions made by the KGE model, we focus on the top ve predictions between genes and pathways in Table 2. By looking at these highly plausible links, we can not only identify novel roles of genes in pathways, but also hypothesize the role of pathways in diseases that has been linked to a given gene.
Gene Database Pathway Score RXRA WikiPathways Nuclear Receptors in Lipid Metabolism and Toxicity 16.20 UPP1 KEGG Pyrimidine metabolism 16.01 EZR KEGG Shigellosis 13.63 UGT1A1 KEGG Porphyrin and chlorophyll metabolism 13.41 BLM WikiPathways DNA IR-damage and cellular response via ATR 12.53 Table 2. Top ten predicted gene-pathway links in which higher scores indicate more plausible links.
The two most con dence predictions suggest that RXRA and UPP1 play
a role in Nuclear Receptors in Lipid Metabolism and Toxicity and Pyrimidine
metabolism pathways, respectively. A survey of the recent biomedical literature
suggests that RXRA is involved in lipid metabolism and UPP1 in the
degradation and salvage of pyrimidine ribonucleosides. Interestingly, the third predicted
link that suggests the involvement of EZR, a cytoplasmic peripheral gene, in the
disease Shigellosis has been previously described by [
We have demonstrated the usage of PyKEEN in predicting missing links in KGs from the biomedical domain. In particular, we performed HPO for three KGE models (i.e., TransE, TransR, and ComplEx) and selected the best performing to provide predictions to a domain expert which manually evaluated the top ranked predictions. Finally, that this work ow that can be applied in any domain highlights the e ectiveness of PyKEEN to discover novel knowledge. Acknowledgement This work was supported by the German national funded BmBF project MLwin.