Semantic Disease Gene Embeddings (SmuDGE): phenotype-based disease gene prioritization without phenotypes

Abstract

Motivation: In the past years, several methods have been developed to incorporate information about phenotypes into computational disease gene prioritization methods. These methods commonly compute the similarity between a disease's (or patient's) phenotypes and a database of gene-to-phenotype associations to find the phenotypically most similar match. A key limitation of these methods is their reliance on knowledge about phenotypes associated with particular genes which is highly incomplete in humans as well as in many model organisms such as the mouse.

Results: We developed SmuDGE, a method that uses feature learning to generate vector-based representations of phenotypes associated with an entity. SmuDGE can be used as a trainable semantic similarity measure to compare two sets of phenotypes (such as between a disease and gene, or a disease and patient). More importantly, SmuDGE can generate phenotype representations for entities that are only indirectly associated with phenotypes through an interaction network; for this purpose, SmuDGE exploits background knowledge in interaction networks comprised of multiple types of interactions. We demonstrate that SmuDGE can match or outperform semantic similarity in phenotype-based disease gene prioritization, and furthermore significantly extends the coverage of phenotype-based methods to all genes in a connected interaction network.

Availability and implementation: https://github.com/bio-ontology-research-group/SmuDGE.

Fig. 1.

Our knowledge graph consists of gene–phenotype associations (encoded using either HPO or MP), disease–phenotype associations (encoded using the HPO), interactions between genes (from the STRING database) and the PhenomeNET ontology

Fig. 2.

Overview over SmuDGE and its applications. (a) On the left side we show the graph of disease–phenotype and gene–phenotype associations together with the PhenomeNET ontology (top), and the same graph including interactions between genes used to generate E-Vecs at the bottom. We generate a corpus by graph traversal and then use a skipgram model to generate vectors for genes and gene products in the graph. These vectors can be used as input to a similarity measure, or a neural network, to predict interactions between genes and diseases. (b) Our ANN model is shown in the center; the input is the pair of disease and gene feature vectors of dimension x, the first hidden layer consists of 2x hidden units, and the second hidden layer consist of x hidden units; we use a dropout of 0.5 to mitigate the effects of overfitting. (c) We evaluate the model by predicting candidate genes for each disease and rank each gene for each disease based on the ANN’s prediction score

Fig. 3.

ROC curves for predicting gene–disease associations using cosine similarity between SmuDGE’s P-Vecs and comparison to Resnik’s semantic similarity measure

Fig. 4.

Comparision of ROC curves for predicting gene–disease associations based on mouse phenotypes using SmuDGE’s feature vectors and comparison to the Resnik and simGIC semantic similarity measures

Fig. 5.

ROC curves for predicting gene–disease associations for diseases with a single or multiple associated genes using SmuDGE’s P-Vec approach

Fig. 6.

ROC cuves for predicting gene–disease associations for diseases with a single or multiple associated genes using SmuDGE’s E-Vec approach