Machine and Deep Learning Methods for Predicting 3D Genome Organization | JoVE Visualize

Area of Science:

Genomics
Computational Biology
Molecular Biology

Background:

Three-dimensional (3D) chromatin interactions, including enhancer-promoter interactions (EPIs), loops, topologically associating domains (TADs), and A/B compartments, are vital for gene expression regulation.
Advanced chromatin conformation capture technologies allow genome-wide 3D structure profiling, even at the single-cell level.
Existing 3D structure catalogs are limited by technological variations, tool differences, and low data resolution, leading to incompleteness and unreliability.

Purpose of the Study:

To review computational tools for predicting three key types of 3D chromatin interactions: EPIs, general chromatin interactions, and TAD boundaries.
To analyze the advantages and disadvantages of existing computational prediction methods.
To identify challenges in computational 3D interaction prediction and propose future research directions.

Main Methods:

The review discusses machine learning approaches that utilize genome annotation data (e.g., ChIP-seq, DNAse-seq), DNA sequencing information (k-mers, transcription factor binding site motifs), and other genomic features.
Methods focus on learning associations between genomic characteristics and 3D chromatin interactions.
The analysis includes a comparative assessment of different computational tools for predicting EPIs, chromatin interactions, and TAD boundaries.

Main Results:

Machine learning offers a promising alternative to address limitations in experimental 3D structure data, enabling the prediction of missing interactions and enhancement of resolution.
Various computational tools exist for predicting EPIs, chromatin interactions, and TAD boundaries, each with specific strengths and weaknesses.
The accuracy and reliability of predictions are influenced by the quality and type of input genomic data and the chosen computational algorithms.

Conclusions:

Computational prediction, particularly using machine learning, is essential for completing and refining 3D chromatin interaction maps.
Further development is needed to overcome obstacles in computational prediction, improving accuracy and reliability.
Future research should focus on developing more robust algorithms and integrating diverse genomic data for comprehensive 3D genome structure analysis.