Contextualizing Genes by Using Text-Mined Co-Occurrence Features for Cancer Gene Panel Discovery

Area of Science:

• Biomedical informatics
• Computational biology
• Genomics

Background:

• Cancer gene panel discovery is crucial for targeted therapies.
• Existing methods may lack comprehensive gene contextualization.
• Biomedical text mining offers a novel approach to literature analysis.

Purpose of the Study:

• To develop and validate a text mining pipeline for cancer gene panel discovery.
• To contextualize genes using text-mined co-occurrence features.
• To assess the pipeline's predictive performance against established gene panels.

Main Methods:

• Applied Biomedical Natural Language Processing (BioNLP) for literature mining.
• Constructed a gene term-feature matrix from scientific literature.
• Validated the pipeline using clinical sequencing data and machine learning models (MSK-IMPACT, Oncomine).

Main Results:

• Achieved 80.8% cosine similarity between text mining and clinical sequencing data for gene frequency.
• Demonstrated high prediction accuracy for gene panels: 0.959 (MSK-IMPACT) and 0.989 (Oncomine).
• Neural network model showed superior prediction performance with an Area Under the ROC Curve (AUC) of 0.992.

Conclusions:

• Text-mined co-occurrence features effectively contextualize genes for cancer discovery.
• The developed pipeline provides a validatable and explainable approach to cancer gene panel identification.
• This method can potentially predict novel cancer-associated genes and refine existing gene panels.