Improving the value of public RNA-seq expression data by phenotype prediction

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Publicly available genomic datasets are crucial for research but often lack essential phenotype annotations.
• Limited phenotype information restricts the usability of genomic data for specific biological questions.

Purpose of the Study:

• To develop and apply an in silico phenotyping approach for predicting missing annotations in genomic data.
• To enhance the utility of large-scale genomic datasets, such as those from the recount2 project.

Main Methods:

• Utilized well-annotated genomic and phenotypic data from consortia like TCGA and GTEx for training.
• Developed predictive models using gene expression data to infer biological phenotypes (sex, tissue, sample source) and experimental conditions (sequencing strategy).
• Applied the in silico phenotyping approach to 70,000 RNA-seq samples within the recount2 project.

Main Results:

• Successfully predicted biological phenotypes and experimental conditions from gene expression data for 70,000 human samples.
• Demonstrated the utility of predicted phenotypes for analyzing public genomic data properties and selecting relevant datasets.
• Showcased the application of predicted phenotypes in downstream genomic analyses.

Conclusions:

• In silico phenotyping significantly improves the annotation and usability of large-scale public genomic data.
• The developed methods and predictions facilitate large-scale expression data analysis, previously not feasible.
• The phenopredict and recount R packages provide accessible tools for phenotype prediction and data utilization.