Deep Learning the Cis-Regulatory Code of Epigenomic Rewiring By Fusion Transcription Factors in Pediatric Leukemia

Pediatric leukemias are commonly driven by chromosomal translocations which create gene fusions involving hematopoietic transcription factors (TFs) in progenitor lymphoid and myeloid cell populations. Although driver gene fusions and altered signaling pathways across leukemia subtypes have been extensively cataloged, the regulatory mechanisms enabling TF fusions to reprogram the epigenome and arrest hematopoietic differentiation remain unclear.

In this study, we generated a single nucleus multiomic atlas from pediatric leukemia patient samples and leveraged deep learning models of regulatory DNA sequence to decipher the regulatory logic linking mutant TFs and regulatory elements to downstream genes and pathways in a sample and cell type-specific manner.

We profiled 22 bone marrow specimens from major diagnostic categories (T-ALL, B-ALL, AML) and including recurrent gene fusions such as ETV6-RUNX1 and RUNX1-RUNX1T1. Using whole genome sequencing (WGS) and multiplexed 10X multiome profiling (single-nucleus RNA+ATAC-seq), we simultaneously profiled gene expression and chromatin accessibility for over 70,000 cells. Unsupervised clustering and cell type annotation revealed 21 distinct clusters, including leukemic and healthy cell populations. Demultiplexing using SNPs from WGS allowed us to recover sample identities and distinguish between malignant and healthy cell populations.

To determine the sequence basis and downstream functional effects of TF rewiring in leukemia, we trained and interpreted ChromBPNet, a fully convolutional neural network, on ATAC-seq data from healthy and malignant B cell clusters. The model discovered enriched motifs for hematopoietic transcription factors, including RUNX1, ETV6, PAX5, and ERG, in ATAC peak regions of ETV6-RUNX1 B-ALL samples.

This integrative approach provides new insights into how oncogenic fusions confer blocked differentiation, survival, and proliferation to leukemia cells. Future work will leverage these models to identify novel motifs for TF fusions and fine-map germline variants to identify functional mutations that perturb TF binding and accessibility through motif disruption.