Deep learning enabled prediction of nuclear lamina-associated chromatin.

Abstract

Tethering of chromatin regions to the nuclear lamina contributes to genome organization and gene regulation. Though several molecular and epigenomic determinants of lamina association have been studied, the influence of genomic sequences has received less attention. Here, we present lamina-associated domain finder (LADDER), a multimodal deep learning model to predict lamina association based on DNA sequence, gene density, and long interspersed nuclear element (LINE)1 and short interspersed nuclear element (SINE) densities. The predictions made by LADDER are cell-type independent. Using LADDER's predictive ability and performing <i>in silico</i> genome-wide random deletions, we established an intrinsic genome lamina susceptibility landscape, where lamina perturbations are enriched near boundary-centered architectural transition zones. Pathogenic structural variants and deletions of CTCF-binding sites alter lamina tethering predominantly when they localize near these structurally sensitive regions. Applying LADDER across species, we examined allele-specific differential lamina association of mouse escapee genes. We further demonstrated the effectiveness of integrating LADDER-predicted nonspecific lamina-tethering data to improve the 3D chromatin models across genomic scales.