Literature mining discerns latent disease-gene relationships

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Gene function dysregulation is a key driver of disease, but mapping gene-pathology relationships is challenging due to genetic complexity and limited computational tools.
• While single-cell gene expression data for healthy tissues is abundant, similar comprehensive data for diseases is lacking, hindering disease mechanism understanding.
• Existing approaches struggle with the diverse genetic manifestations of diseases and confounding clinical factors.

Purpose of the Study:

• To develop a computational method for identifying gene-disease associations by analyzing scientific literature.
• To leverage natural language processing and deep learning to predict novel gene-pathology relationships.
• To create a resource that aids in understanding the genetic basis of diseases.

Main Methods:

• Mined approximately 18 million PubMed abstracts, selecting 4.5 million related to gene roles in disease pathogenesis.
• Fine-tuned a pre-trained Bidirectional Encoder Representations from Transformers (BERT) model for biological language modeling.
• Trained the model to learn vector representations of biological entities (genes, diseases, cell types) preserving their relationships.

Main Results:

• The fine-tuned BERT model, PathoBERT, successfully predicted disease-gene associations not present in the training data.
• Demonstrated the feasibility of in silico hypothesis generation for linking biological entities.
• The model effectively captures complex relationships between genes and diseases from unstructured text.

Conclusions:

• PathoBERT offers a powerful new tool for discovering gene-disease associations.
• This approach facilitates the in silico synthesis of biological hypotheses, accelerating research into disease mechanisms.
• The developed model and associated resources can significantly advance the study of genetic pathology.