InsuLock: A Weakly Supervised Learning Approach for Accurate Insulator Prediction, and Variant Impact Quantification.

Area of Science:

• Genomics
• Computational Biology
• Epigenetics

Background:

• Chromatin insulators are vital for genome regulation, but current mapping methods are costly and imprecise.
• Accurate insulator identification is key for understanding genome evolution, biological functions, and disease variant impacts.

Purpose of the Study:

• To introduce InsuLock, a novel weakly supervised deep learning approach for efficient and high-resolution chromatin insulator mapping.
• To enable precise localization of insulator boundaries and accurate quantification of variant effects on insulator function.

Main Methods:

• InsuLock employs a Siamese neural network for insulator presence prediction.
• An object detection module with gradient-weighted class activation mapping achieves ~40 bp resolution for boundary localization.
• Variant impact is assessed by comparing insulator scores between wild-type and mutant alleles.

Main Results:

• InsuLock demonstrated superior performance over existing methods with an AUROC of ~0.96.
• The method significantly condensed insulator annotations (to ~2.5% of original size) while improving conservation and motif enrichment.
• Cell-type-specific variant impacts were analyzed using brain scATAC-seq data.

Conclusions:

• InsuLock offers a powerful, cost-effective solution for high-resolution chromatin insulator mapping.
• The tool accurately quantifies variant impacts and aids in identifying disease-associated genetic elements.
• A schizophrenia GWAS variant was identified disrupting an insulator loop, suggesting a novel disease mechanism.