GhostBuster: A Deep-Learning-based, Literature-Unbiased Gene Prioritization Tool for Gene Annotation Prediction

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• A significant number of human protein-coding genes are poorly characterized, referred to as "ghost genes".
• Research literature exhibits a "bandwagon effect", disproportionately focusing on well-annotated genes, which introduces bias.
• This literature bias influences machine learning (ML) models, leading to predictions that favor well-studied genes and potentially overestimate biological relevance.

Purpose of the Study:

• To develop a machine learning (ML) platform, GhostBuster, designed to predict gene functions, disease associations, and interactions while minimizing literature bias.
• To evaluate the impact of biased (Gene Ontology) versus unbiased training datasets (LINCS, TCGA, STRING) on ML model performance and bias amplification.

Main Methods:

• Developed GhostBuster, an encoder-decoder ML platform.
• Compared ML models trained on literature-biased datasets against those trained on unbiased datasets (LINCS, TCGA, STRING).
• Assessed the models' effectiveness in identifying novel gene annotations and predicting gene functions, disease associations, and interactions.

Main Results:

• Literature-biased datasets yielded higher ML metrics but amplified existing biases.
• Models trained on unbiased datasets were 2-3 times more effective at identifying recently discovered gene annotations.
• The TCGA dataset, with minimal literature bias, demonstrated robust performance (ROC-AUC of 0.8-0.95).

Conclusions:

• GhostBuster is the first ML framework explicitly designed to counteract literature bias in gene annotation.
• The platform can predict novel gene functions, refine pathway memberships, and prioritize intergenic GWAS hits.
• GhostBuster provides a powerful tool for exploring the roles of understudied genes in cellular function, disease, and molecular networks.