Mimvec: a deep learning approach for analyzing the human phenome

Area of Science:

• Genomics
• Bioinformatics
• Natural Language Processing

Background:

• Traditional methods for disease gene inference often use TF-IDF, which ignores semantic relationships and creates high-dimensional vectors.
• Existing frameworks struggle to capture the intrinsic semantic characteristics of biomedical documents.

Purpose of the Study:

• To propose mimvec, a novel framework utilizing deep learning for human phenome analysis.
• To overcome the limitations of traditional TF-IDF methods in capturing semantic nuances in biomedical data.

Main Methods:

• Developed mimvec, a deep learning approach for analyzing the human phenome.
• Converted 24,061 Online Mendelian Inheritance in Man (OMIM) records into low-dimensional vectors.
• Derived pairwise phenotype similarities for 7988 human inherited diseases.

Main Results:

• Vector representations enabled effective classification of phenotype and gene records.
• Successfully discriminated diseases based on inheritance styles and mechanisms.
• Demonstrated that phenotype overlap implies genotype overlap when analyzing phenome data with genomic data.
• Prioritized candidate genes using derived phenotype similarities, showing advantages over existing methods.

Conclusions:

• The mimvec method captures semantic relationships and mitigates the dimensionality issues of TF-IDF.
• This approach is expected to have wide applications in analyzing the growing volume of electronic health records for precision medicine.