Deep semi-supervised learning ensemble framework for classifying co-mentions of human proteins and phenotypes

Area of Science:

• Bioinformatics
• Biomedical Natural Language Processing
• Computational Biology

Background:

• Accurate identification of human protein-phenotype relationships is crucial for understanding rare and complex diseases.
• Experimental validation of these associations is costly and time-consuming, driving demand for automated extraction tools.
• Manual annotation for training models is resource-intensive, while unlabeled data is abundant.

Purpose of the Study:

• To propose a novel deep semi-supervised ensemble framework for classifying human protein-phenotype co-mentions.
• To leverage large amounts of unlabeled data alongside a small labeled dataset to improve model performance.
• To develop a prototype system, PPPredSS, for practical application.

Main Methods:

• A deep semi-supervised ensemble framework combining deep neural networks, semi-supervised learning, and ensemble learning.
• Integration of sophisticated language models, convolutional neural networks, and recurrent neural networks within the PPPredSS prototype.
• Utilizing both labeled and extensive unlabeled sentence-level co-mentions of human proteins and phenotypes.

Main Results:

• The proposed framework achieved state-of-the-art performance in classifying human protein-phenotype co-mentions.
• PPPredSS outperformed existing supervised and semi-supervised methods.
• Case studies demonstrated the system's utility as a curation assistant for biologists.

Conclusions:

• A novel deep, semi-supervised, and ensemble learning approach for human protein-phenotype co-mention classification was presented.
• The findings offer significant implications for biomedical researchers, biocurators, and the text mining community.
• This work advances biomedical relationship extraction methodologies.