Application of BERT to Enable Gene Classification Based on Clinical Evidence

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Accurate classification of cancer-related genes is crucial for diagnosis and treatment.
• Current manual classification of genetic mutations is subjective, time-consuming, and pathologist-dependent.
• Computational approaches are emerging for automated mutation analysis, but face challenges like text complexity and inconsistent interpretation.

Purpose of the Study:

• To adapt a deep learning method, Bidirectional Encoder Representations from Transformers (BERT), for automated classification of genetic mutations using text evidence.
• To address challenges in automated mutation classification, including text length, data bias, and redundancy.

Main Methods:

• Utilized an annotated database of genetic mutations and associated text.
• Applied the Bidirectional Encoder Representations from Transformers (BERT) deep learning model for text classification.
• Trained the BERT model, specifically addressing issues of extreme text length, biased data, and high repeatability.

Main Results:

• The BERT+abstract model achieved a logarithmic loss of 0.80, recall of 0.6837, and F-measure of 0.705.
• Demonstrated the feasibility of using BERT for classifying genomic mutation text from literature-based datasets.
• Indicated satisfactory performance in automating the classification of genetic mutations.

Conclusions:

• Bidirectional Encoder Representations from Transformers (BERT) is a practical and effective tool for classifying genomic mutation text.
• This deep learning approach can significantly accelerate cancer research, aiding in tumor progression understanding, diagnosis, and treatment design.
• Automating genetic mutation classification enhances the accuracy and efficiency of clinical interpretation in oncology.