Unsupervised Event Graph Representation and Similarity Learning on Biomedical Literature.

Area of Science:

• Bioinformatics
• Computational Biology
• Natural Language Processing

Background:

• Automatic extraction of biomedical events from literature is crucial for understanding complex biological interactions.
• Existing methods lack effective approaches for learning embeddings or similarity metrics for event graphs, hindering machine learning applications.
• This gap limits the discovery of unlinked biological relations and the advancement of data-driven biological research.

Purpose of the Study:

• To propose Deep Divergence Event Graph Kernels (DDEGK), an unsupervised inductive method for mapping biomedical events into low-dimensional vectors.
• To preserve both structural and semantic similarities of event graphs without requiring task-specific labels or feature engineering.
• To enable machine learning techniques for promoting discoveries in biological relations.

Main Methods:

• DDEGK utilizes deep graph kernel solutions and pre-trained language models.
• It employs cross-graph attention networks for pairwise alignment and transformer-based models for encoding attributes.
• The method compares events against anchor events, operating at the graph level.

Main Results:

• Learned event representations effectively support graph classification, clustering, and visualization tasks.
• The method facilitates downstream semantic textual similarity analysis.
• DDEGK significantly outperforms existing state-of-the-art methods across nine biomedical datasets.

Conclusions:

• DDEGK provides a novel unsupervised approach for learning meaningful representations of biomedical events.
• The method bridges the gap in event graph analysis, enabling advanced machine learning applications in biology.
• DDEGK demonstrates superior performance and broad applicability in various bioinformatics tasks.