Statistical principle-based approach for gene and protein related object recognition

Area of Science:

• Biomedical text mining
• Bioinformatics
• Natural Language Processing (NLP)

Background:

• Extracting valuable information like genes and proteins from chemical and pharmaceutical patents is crucial for biomedical research.
• The BioCreative V.5 challenge focused on gene- and protein-related object (GPRO) recognition in patents, including linking mentions to database records.

Purpose of the Study:

• To develop and evaluate a system for identifying gene and protein mentions in patents.
• To assess the system's performance in the BioCreative V.5 GPRO recognition task.

Main Methods:

• The SPBA-CRF system integrates a statistical-principle-based approach (SPBA) for gene mention recognition with conditional random fields (CRF).
• SPBA predictions were used as features for the CRF-based recognizer, originally developed for chemical mention identification and adapted for GPRO recognition.

Main Results:

• The SPBA-CRF system achieved an F-score of 73.73% for GPRO type 1 and 78.66% for combined GPRO types 1 and 2 in the BioCreative V.5 task.
• SPBA, trained on an external NER dataset, demonstrated good performance on partial match evaluation.
• SPBA significantly enhanced the performance of the CRF-based recognizer trained on the GPRO dataset.

Conclusions:

• The SPBA-CRF system effectively identifies gene and protein-related objects in patent literature.
• The study highlights the utility of SPBA in improving GPRO recognition accuracy and its robustness with external datasets.