MIRACN: a residual convolutional neural network for predicting cell line specific functional regulatory variants

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Interpreting noncoding genetic variants post-genome-wide association studies presents a significant challenge due to their complexity and poorly understood functions.
• Understanding the functional impact of these variants is crucial for advancing precision medicine and disease research.

Purpose of the Study:

• To develop a novel computational tool, MIRACN, for predicting cell line-specific functional regulatory variants.
• To enhance the understanding of noncoding variant functionality and their context-specific regulatory mechanisms.

Main Methods:

• Developed MIRACN, a residual convolutional neural network model.
• Utilized a large dataset from massively parallel reporter assays (MPRAs).
• Employed a multitask learning strategy trained across seven distinct cell lines.

Main Results:

• MIRACN achieved superior performance in predicting cell line-specific regulatory variants compared to existing methods.
• Demonstrated high accuracy in identifying functional variants and their cell-specific regulatory mechanisms on an independent MPRA test dataset.
• MIRACN provides functional scores and pinpoints the specific cell line where variants exhibit function.

Conclusions:

• MIRACN significantly improves the resolution of research into noncoding variant functionality.
• The tool offers valuable insights into cellular context-specific regulatory mechanisms.
• This advancement facilitates more precise diagnostic and therapeutic strategies for genetic diseases.