An end-to-end generalizable deep learning framework to comprehensively analyze transcriptional regulation

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Genome annotation is crucial for understanding biological functions.
• Current methods involve expensive and time-consuming molecular assays across numerous samples.
• This limits the scope of annotation across diverse species and conditions.

Purpose of the Study:

• To introduce BioSeq2Seq, a novel deep learning framework for efficient and cost-effective genome annotation.
• To enable flexible genome annotation tasks using parameterized configurations and architectural refinement.
• To provide an alternative to traditional molecular assays for inferring cell-line-specific genomic features.

Main Methods:

• Utilized a tri-modal input: conserved DNA sequence features, cell-line-specific transcriptional activity, and directionality from a single run-on sequencing assay.
• Developed a deep learning framework (BioSeq2Seq) capable of inferring molecular assays for genome annotation.
• Employed parameterized configurations and gradient-guided architectural refinement for task-specific optimization.

Main Results:

• Achieved high accuracy across four downstream genome annotation tasks.
• Demonstrated significant improvements in histone modification prediction (14.27%), functional element identification (2.50%), and gene expression prediction (2.90%).
• Maintained performance comparable to state-of-the-art methods in transcription factor binding site (TFBS) prediction.

Conclusions:

• BioSeq2Seq offers an efficient and low-cost solution for genome annotation.
• The framework provides competitive performance using single-cell-line input data.
• Enables broader and more accessible genome annotation across various biological contexts.