Predicting 3D genome folding from DNA sequence with Akita.

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Molecular Biology

Background:

• The human genome folds into specific 3D structures in interphase, influencing gene regulation.
• Cohesin and CTCF (CCCTC-binding factor) are crucial for genome folding, but the sequence-level determinants remain unclear.

Purpose of the Study:

• To develop a predictive model for locus-specific genome folding directly from DNA sequence.
• To understand the sequence grammar underlying genome structure and function.

Main Methods:

• A convolutional neural network, named Akita, was designed to predict 3D genome folding from DNA sequence.
• The model was trained on existing genome folding data, learning nucleotide-level features.

Main Results:

• Akita accurately predicts genome folding patterns using only DNA sequence information.
• The model identified an orientation-specific grammar for CTCF binding sites and nucleotide effects beyond the core motif.
• Akita enables rapid in silico predictions, including saturation mutagenesis and eQTL interpretation.

Conclusions:

• DNA sequence alone contains sufficient information to predict 3D genome folding.
• Akita provides a powerful tool for decoding genome function through its structural organization.
• The findings offer new avenues for studying genome regulation, structural variants, and species-specific folding.