Evaluation of input data modality choices on functional gene embeddings.

Area of Science:

• Genomics and Bioinformatics
• Machine Learning in Biology

Background:

• Functional gene embeddings represent gene function numerically for machine learning.
• Self-supervised learning algorithms create these embeddings from diverse data types like omics, networks, and literature.
• Lack of comparative studies on data modalities for embedding construction hinders optimal model development.

Purpose of the Study:

• To benchmark functional gene embeddings derived from different data modalities.
• To evaluate their performance in predicting disease-gene associations, cancer drivers, and phenotype-gene links.
• To assess utility for genome-wide association study (GWAS) signals.

Main Methods:

• Generated functional gene embeddings using various data sources (omics, PPI networks, literature, protein sequences).
• Benchmarked embeddings by training off-the-shelf predictors for downstream tasks.
• Compared performance against dedicated state-of-the-art predictors.

Main Results:

• Precomputed embeddings with simple predictors matched or surpassed specialized models.
• Literature and low-throughput PPI-based embeddings excelled in predicting curated gene lists.
• These embeddings showed bias towards well-studied genes and did not improve GWAS signal prediction.

Conclusions:

• Functional gene embeddings are highly useful for machine learning in genetics.
• Embeddings from literature and low-throughput PPIs are effective but potentially biased.
• Careful consideration of data source biases is crucial for reliable machine learning applications in genetics.