Text mining-based word representations for biomedical data analysis and protein-protein interaction networks in...

Area of Science:

• Computational Biology
• Bioinformatics
• Natural Language Processing

Background:

• Biomedical literature is rapidly expanding, creating a need for automated knowledge extraction.
• Discovering relationships between biological entities is crucial for advancing scientific understanding.
• Existing methods for knowledge extraction from text require enhancement.

Purpose of the Study:

• To generate word vector representations (embeddings) from a large biomedical text corpus using the word2vec approach.
• To develop and validate a text mining pipeline for creating biomedical word embeddings.
• To assess the utility of these embeddings in constructing biological networks and aiding in cancer data analysis.

Main Methods:

• Utilized the word2vec model on over 16 million PubMed abstracts.
• Implemented a text mining pipeline with pre-processing steps, including synonym substitution.
• Extracted gene-gene networks from word embeddings to train Graph-Convolutional Neural Networks (CNNs).
• Compared performance against networks derived from protein-protein interactions (PPI) and other embedding algorithms.

Main Results:

• Word embeddings captured biologically meaningful relationships, evidenced by high cosine similarity with known biological interactions (PPIs, pathways, functions).
• Graph-CNNs trained with word2vec-derived networks showed competitive performance in metastatic event prediction tasks.
• The size of the corpus influenced the variability of word representations.

Conclusions:

• Word embeddings generated via text mining are effective for capturing biologically relevant entity relationships.
• The developed word embeddings provide a valuable resource for constructing biological networks and advancing biomedical research.
• A web service was created for exploring relationships between biomedical terms using these annotated embeddings.