Figure 1.

Overview of CITRUS: An attention-based model with TF-target gene priors. The input to our framework includes somatic alteration and copy number variation, assay for transposase-accessible chromatin with high-throughput sequencing (ATAC-seq), tumor expression datasets and TF recognition motifs. CITRUS takes somatic alteration and copy number variation data as input and encodes them as a tumor embedding using a self-attention mechanism. Cancer type embedding is used to stratify the confounding factor of tissue type. The middle layer further transforms the tumor embeddings into a TF layer, which represents the inferred activities of 320 TFs. Finally, gene expression levels are predicted from the TF activities through a TF-target gene priors constrained sparse layer based on ATAC-seq.