BioByGANS

BioByGANS integrates contextual and syntactic features using a graph attention network to perform biomedical named entity recognition by classifying words as nodes in a sentence-level graph for entity extraction from biomedical text.

Key Features:

• Graph-Based Modeling: Formulates BioNER as a node classification problem within a sentence-level graph that represents dependencies and topology.
• Contextual and Syntactic Feature Integration: Encodes contextual features with BioBERT and preprocesses syntactic features (part-of-speech tags, dependencies, sentence topology) with SpaCy to create dual feature representations.
• Graph Attention Network (GAT): Applies a graph attention network to fuse contextual and syntactic encodings and to emphasize important node relationships within the graph.
• Softmax Classification: Uses a softmax classifier to compute class probabilities for node (word) entity labels.
• Sentence and Word Segmentation: Segments sentences by periods and words by spaces and symbols as stated.
• Benchmark Performance: Evaluated on BC2GM, JNLPBA, BC4CHEMD, BC5CDR-chem, BC5CDR-disease, NCBI-disease, Species-800, and LINNAEUS with reported F1-scores from 75.01% to 94.74%, reported to outperform existing state-of-the-art BioNER methods.

Scientific Applications:

• Structured data extraction: Converts unstructured biomedical literature into structured entity data for downstream analyses.
• Knowledge extraction: Facilitates extraction of biomedical knowledge from text corpora.
• Drug discovery support: Provides entity recognition that can aid data curation and information retrieval for drug discovery research.
• Disease gene identification: Assists identification and curation of disease-related genetic entities.
• Biological pathway analysis: Supports extraction of entities relevant to biological pathway reconstruction and analysis.
• Large-scale text corpora: Applicable to diverse, large-scale biomedical text corpora and benchmark BioNER datasets.

Methodology:

Sentences are segmented by periods and words by spaces and symbols; contextual features are encoded with BioBERT and syntactic features are preprocessed with SpaCy; a graph representing sentence dependencies and topology is constructed, a graph attention network integrates the encoded features, and a softmax function is used for final entity classification.