Benchmarking knowledge graph embedding models for the prediction of oligogenic combinations.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Identifying oligogenic causes of rare diseases is a persistent challenge due to limited training data and difficulties in feature selection for predictive models.
• Existing predictive and ranking approaches show limited precision in identifying pathogenic gene pairs for complex genetic disorders.

Purpose of the Study:

• To investigate the utility of structured biological information integrated into a heterogeneous knowledge graph for learning genetic representations.
• To benchmark state-of-the-art knowledge graph embedding (KGE) models for identifying potentially pathogenic gene pairs involved in oligogenic diseases.

Main Methods:

• An exhaustive benchmarking of various KGE models was performed to assess their performance in predicting pathogenic gene pairs.
• Models were evaluated using cross-validation, a holdout set, and a cohort of new male infertility cases.
• Specific attention was paid to preventing data leakage in the embedding space during cross-validation.

Main Results:

• KGE models achieved high accuracy in predicting pathogenic gene pairs, with an Area Under the Precision-Recall Curve reaching up to 0.93.
• The performance represents a significant advancement over previous methods for predicting gene pairs in oligogenic diseases.
• Translational Distance models (TransE, MurE, RotatE) and Semantic Matching models (DisMult, QuatE) demonstrated superior performance.

Conclusions:

• Structured biological information and KGE methods offer a powerful approach for advancing the prediction of gene pairs implicated in oligogenic rare diseases.
• Careful cross-validation is crucial to avoid overly optimistic results due to data leakage in the embedding space.
• Future work is needed to develop methods that provide explanations for the identified gene pair relevance in oligogenic diseases.