Estimating genotype-tissue specific gene expression using hybrid deep learning

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Genotype-tissue expression (GTE) profiles are crucial for understanding genetic variation's impact on gene regulation.
• Existing GTE datasets are often incomplete, and experimental profiling is resource-intensive.
• Current computational methods lack the integration of genomic context and neighboring gene expression for imputation.

Purpose of the Study:

• To develop a novel computational model for accurate multi-tissue GTE profile imputation.
• To address limitations of existing methods by incorporating genomic context and expression data from neighboring genes.
• To provide a scalable and cost-effective alternative to experimental GTE profiling.

Main Methods:

• A hybrid deep learning model integrating a convolutional neural network (CNN), transformer encoder, and XGBoost regressor was developed.
• The model utilizes promoter sequences, tissue correlations, intergene distances, and gene orientation.
• The model was validated by completing missing profiles in the Genotype-Tissue Expression (GTEx) dataset.

Main Results:

• The novel model achieved approximately 30% higher accuracy compared to distance-based imputation methods.
• Generated expression profiles demonstrated strong alignment with experimental data.
• Successfully imputed missing GTE profiles within the GTEx dataset, showcasing practical utility.

Conclusions:

• The developed hybrid deep learning model offers a highly accurate and efficient solution for estimating GTE profiles.
• This approach enables cost-effective GTE profile generation, especially for low-expression genes and under-characterized genomes.
• The model facilitates advancements in genomics research by overcoming experimental data scarcity.