Predicting implicit concept embeddings for singular relationship discovery replication of closed literature-based...

Area of Science:

• Computational Biology
• Bioinformatics
• Artificial Intelligence in Medicine

Background:

• Literature-based discovery (LBD) leverages existing literature to uncover new knowledge by connecting implicit relationships.
• Current LBD systems increasingly use deep learning (DL) and knowledge graphs (KGs), often framing discovery as knowledge graph completion (KGC).
• Existing KGC-based LBD methods require domain expertise for effective query submission.

Purpose of the Study:

• To develop a novel LBD approach that identifies all implicit hypotheses from a given query, expediting knowledge discovery.
• To revise the KGC task for predicting interconnecting vertex embeddings within a KG.
• To reduce researcher burden in hypothesis generation.

Main Methods:

• A similarity learning objective was used to train a model for predicting linking vertex embeddings.
• Model predictions were compared against known vertices to determine the likelihood of implicit relationships.
• Exploration of edge representation methods (average, concatenation, Hadamard) and input representation scaling for faster convergence.

Main Results:

• The method successfully replicated five known discoveries from the Hallmark of Cancer (HOC) dataset.
• Performance was comparable or optimal when compared to two existing LBD works across five datasets.
• Statistical significance testing confirmed the method's efficacy in predicting linking vertex embeddings.

Conclusions:

• The similarity-based learning objective effectively predicts linking vertex embeddings for hypothesis discovery.
• The approach provides a ranked list of potential links, reducing researcher effort.
• This facilitates further exploration of generated hypotheses in scientific research.