IceQream: Quantitative chromosome accessibility analysis using physical TF models

Single-cell mapping of chromosomal accessibility patterns has recently led to improved predictive modelling of epigenomic activity from sequence. However, quantitative models explaining the epigenome using directly interpretable components are still lacking. Here we develop IceQream (IQ), a modelling strategy and inference algorithm for regressing accessibility from sequences using physical models of transcription factor (TF) binding. IQ uses spatial integration of sequences over a range of TF-DNA affinities and localization relative to the target locus. It infers TF effective concentrations as latent variables that activate or repress regulatory elements in a non-linear fashion. These are supplemented with synergistic and antagonistic pairwise interactions between TFs. Analysis of both human and mouse data shows that IQ derives similar, and in some cases, better performance compared to state-of-the-art deep neural network models. IQ provides an essential mechanistic and explicable baseline for further developments toward understanding gene and genome regulation from sequence. Overall design: Single nuclei were extracted from mouse embryos harvested from timed pregnant Hsd:ICR (CD-1) females at embryonic days E6.5-E7.5. Nuclei isolation was performed using a modified low cell input protocol with lysis buffer containing 10 mM Tris-HCl (pH7.4), 10 mM NaCl, 3 mM MgCl2, 0.1% Tween-20, 0.1% Nonident P40 substitute, 0.01% digitonin, 1% BSA, 1 mM DTT, and RNase inhibitor. Single-cell ATAC and gene expression libraries were prepared using the 10X Genomics Multiome platform with Chromium Next GEM Chip J and sequenced on an Illumina NovaSeq 6000 instrument.