Enhancing Gene Expression Predictions Using Deep Learning and Functional Annotations.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Transcriptome-wide association studies (TWAS) link genetic variations to traits via gene expression.
• Current TWAS methods rely on linear models (e.g., elastic net) for gene expression prediction, which miss nonlinear genetic effects.
• Accurate gene expression prediction is vital for the power of TWAS.

Purpose of the Study:

• To develop a deep learning model for improved gene expression prediction from genotype data.
• To capture nonlinear relationships and higher-order interactions missed by linear models.
• To leverage functional annotations to enhance prediction accuracy in TWAS.

Main Methods:

• Proposed a deep learning architecture with a learnable input scaling layer and convolutional encoder.
• Implemented parameter sharing across networks to incorporate functional annotation information.
• Evaluated the model against elastic net regression using real-world genomic datasets.

Main Results:

• The deep learning model consistently outperformed elastic net regression in predicting gene expression for heritable genes.
• Leveraging functional annotations significantly improved the deep learning model's predictive performance.
• Elastic net regression did not achieve similar performance gains when using functional annotations.

Conclusions:

• The proposed deep learning method effectively models nonlinear genetic effects for gene expression prediction.
• This approach offers superior performance over linear models in TWAS, particularly when incorporating functional genomic data.
• The method enhances the ability to uncover genotype-phenotype relationships through more accurate gene expression imputation.