ºÚÁÏÍø

Multimodal Regulatory Genomics with Deep Learning — Case Study in ccRCC

Author: Zahra Dashtgerd
Supervisor: Juho Rousu

Abstract: Chromatin accessibility profiling by ATAC-seq reveals the regulatory landscape of non-coding DNA, yet decoding the sequence-level grammar that drives cell-type-specific regulatory programs remains challenging. This thesis investigates whether deep learning can model and interpret chromatin accessibility directly from DNA sequence in clear cell renal cell carcinoma (ccRCC). ChromBPNet models were trained independently on ATAC-seq data from two renal epithelial cell lines: HK2, representing normal proximal tubule epithelium, and RCC-JF, a VHL-mutant ccRCC line. Both models achieved Pearson R values of 0.740 and 0.716 on held-out chromosomes, matching published benchmarks and confirming reliable sequence-to-accessibility prediction. Contribution scores computed via DeepLIFT and de novo motif discovery with TF-MoDISco recovered transcription factor families consistent with known renal and ccRCC biology, including PAX, HNF1B, ETS, AP-1, ATF, and TEAD. Comparative analysis of PAX co-factor activity across 7,596 shared accessible regions revealed cell-type-specific regulatory rewiring: PAX–HNF1B co-occurrence was enriched in RCC-JF, consistent with their known cooperative role in ccRCC; HNF4 motifs were present genome-wide in RCC-JF but absent in HK2, reflecting retained proximal tubule lineage identity; and ZNF motifs dominated the HK2 regulatory landscape but were nearly absent in RCC-JF, suggesting loss of a repressive program upon malignant transformation. These results demonstrate that ChromBPNet captures biologically interpretable, cell-type-specific regulatory logic from sequence alone, offering a scalable framework for variant effect prediction and in silico mutagenesis in regulatory genomics.