A general framework for predicting the transcriptomic consequences of non-coding variation and small molecules | JoVE Visualize

Area of Science:

Genomics
Computational Biology
Bioinformatics

Background:

Genome-wide association studies (GWASs) identify numerous genetic loci for complex traits, but most variants are in non-coding regions, hindering functional interpretation.
Translating GWAS findings into biological understanding is challenging due to the difficulty in pinpointing the functional impact of non-coding variants.

Purpose of the Study:

To develop a computational framework, peaBrain, that leverages convolutional neural networks to model tissue-specific gene expression and interpret the functional consequences of genetic variation.
To improve the prediction of gene abundance and identify functional non-coding variants associated with complex traits and diseases.

Main Methods:

Developed peaBrain, a two-stage convolutional neural network model predicting gene abundance based on tissue-specific transcriptional machinery and genotype variation.
Calculated non-coding impact scores, correlated them with evolutionary constraint, and assessed their predictive power for disease association and transcription factor binding.
Evaluated peaBrain's ability to pinpoint functional tissues underlying complex traits and predict transcriptomic impacts of small molecules and shRNA.

Main Results:

peaBrain explains over 50% of the variance in mean transcript abundance across most tissues and outperforms linear models in predicting genotype-specific expression.
Non-coding impact scores derived from peaBrain correlate with evolutionary constraint and predict disease-associated variants and allele-specific transcription factor binding.
peaBrain effectively identifies functional tissues for complex traits, outperforming colocalization methods, and predicts transcriptomic impacts comparable to experimental validation.

Conclusions:

peaBrain offers a powerful computational approach to interpret GWAS findings, particularly non-coding variants, by modeling gene expression dynamics.
The framework enhances the identification of functional genetic elements, aids in understanding complex trait genetics, and has broad applications in drug discovery and biological process modeling.