A Gene Set Foundation Model Pre-Trained on a Massive Collection of Diverse Gene Sets

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Foundation models excel at pattern recognition in large datasets, generating embeddings for diverse applications.
• Gene function prediction is crucial for understanding biological systems and disease mechanisms.

Purpose of the Study:

• To develop and benchmark a gene set foundation model (GSFM) for accurate gene function prediction.
• To compare the GSFM's performance against existing state-of-the-art methods.
• To systematically predict functions for all human genes using the best-performing GSFM architecture.

Main Methods:

• Trained a GSFM on a large corpus of unlabeled gene sets from Rummagene and RummaGEO databases.
• Benchmarked various foundation model architectures and training data sources for gene function prediction.
• Compared GSFM predictions with established gene function prediction tools.

Main Results:

• One GSFM architecture demonstrated superior performance over all other evaluated methods and models.
• The optimized GSFM was utilized for genome-wide prediction of human gene functions.
• Predicted gene functions are accessible via web-based gene pages.

Conclusions:

• The developed GSFM offers a powerful and accurate approach for gene function prediction.
• This work provides a valuable resource for researchers by systematically annotating human gene functions.
• The GSFM has the potential to advance various areas of biological research and drug discovery.