Mining influential genes based on deep learning.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Large-scale gene expression profiling aids in understanding disease, genetic perturbation, and drug action.
• The L1000 profiling method offers a low-cost, high-throughput approach but relies on landmark genes with suboptimal predictive power.
• Improved computational methods are needed to identify more informative landmark genes for deeper genomic analysis.

Purpose of the Study:

• To develop a computational framework for identifying a more informative subset of landmark genes.
• To enhance the accuracy and robustness of inferring genomic information from reduced gene expression profiles.
• To improve the analysis of biological big data for discovering functional connections.

Main Methods:

• A deep learning framework utilizing an AutoEncoder to model non-linear gene relationships.
• Application of DeepLIFT to compute gene importance scores for data-driven landmark gene selection.
• Validation of the new landmark gene set using mean absolute error (MAE) and Pearson correlation coefficient (PCC).

Main Results:

• A novel set of landmark genes was identified using the proposed deep learning framework.
• The new landmark genes demonstrated superior accuracy and robustness in predicting target genes compared to the L1000 set.
• The identified landmark genes capture more comprehensive genomic information.

Conclusions:

• The proposed framework is effective for analyzing biological big data and uncovering biological insights.
• The newly identified landmark genes can improve gene expression profiling for research into functional connections.
• This approach facilitates a deeper understanding of complex biological systems through enhanced gene expression analysis.