Improved prediction of smoking status via isoform-aware RNA-seq deep learning models

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Gene splicing is crucial for eukaryotic gene expression but often overlooked in predictive models.
• Cigarette smoking significantly impacts gene expression and is a major risk factor for various diseases.

Purpose of the Study:

• To develop deep neural network models predicting smoking status using RNA sequencing data.
• To investigate the utility of exon and isoform level gene expression data, beyond gene-level data, for predictive modeling.

Main Methods:

• Utilized RNA sequencing data (gene, exon, and isoform quantifications) from 2,557 subjects in the COPDGene Study.
• Developed deep neural network models with a focus on incorporating exon-to-isoform mapping.
• Compared the predictive performance of gene-level, exon-level, and isoform-level models.

Main Results:

• Exon and isoform quantifications significantly outperformed gene-level models in predicting smoking status.
• Exon-based models achieved an Area Under the Receiver Operating Characteristic curve (AUC) of 0.88, improving to 0.94 with a larger gene set.
• The developed models demonstrated superior performance compared to a previously published gene-level model (AUC 0.82).

Conclusions:

• Incorporating gene splicing information, specifically at the exon and isoform levels, substantially enhances the accuracy of clinical prediction models.
• Isoform variability in RNA-seq data represents a valuable, underutilized source of information for improving predictive capabilities.
• Deep learning approaches leveraging detailed RNA-seq data can lead to more robust biomarkers for environmental exposures like smoking.