Gene expression inference with deep learning.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Large-scale gene expression profiling is crucial for understanding cellular states but remains expensive.
• Current methods like linear regression (LR) struggle with complex, nonlinear gene expression relationships.
• The NIH LINCS program uses landmark gene profiling to reduce costs, but LR's accuracy is limited.

Purpose of the Study:

• To develop a deep learning method (D-GEX) for inferring target gene expression from landmark gene expression.
• To improve the accuracy of gene expression inference compared to existing computational methods.
• To provide a cost-effective strategy for generating large-scale gene expression compendia.

Main Methods:

• Developed D-GEX, a deep learning model for gene expression inference.
• Trained D-GEX on the Gene Expression Omnibus (GEO) dataset (111K microarray profiles).
• Validated D-GEX on an independent RNA-Seq-based GTEx dataset (2921 profiles).

Main Results:

• Deep learning (D-GEX) significantly outperformed linear regression (LR) with a 15.33% relative improvement in mean absolute error on the GEO dataset.
• D-GEX achieved lower error than LR for 99.97% of target genes in the GEO dataset.
• On the GTEx dataset, D-GEX showed a 6.57% relative improvement over LR, with lower error for 81.31% of target genes.

Conclusions:

• Deep learning offers a more accurate and cost-effective approach for inferring gene expression compared to linear regression.
• D-GEX provides a powerful tool for analyzing large-scale gene expression data, advancing cellular state characterization.
• The D-GEX method is publicly available, facilitating its adoption in genomic research.